CN106447721A

CN106447721A - Image shadow detection method and device

Info

Publication number: CN106447721A
Application number: CN201610817703.2A
Authority: CN
Inventors: 姚聪; 周舒畅; 周昕宇; 何蔚然; 印奇
Original assignee: Beijing Megvii Technology Co Ltd; Beijing Aperture Science and Technology Ltd
Current assignee: Beijing Megvii Technology Co Ltd; Beijing Aperture Science and Technology Ltd
Priority date: 2016-09-12
Filing date: 2016-09-12
Publication date: 2017-02-22
Anticipated expiration: 2036-09-12
Also published as: CN106447721B

Abstract

The embodiment of the invention provides an image shadow detection method and device. The method comprises the steps: obtaining a to-be-detected image; carrying out the to-be-detected image through a full-convolution network, so as to obtain a detection result about the position information of the shadow region in the to-be-detected image. According to the embodiment of the invention, the method and device can effectively detect the shadow region in the detected image at high precision through the full-convolution network, thereby facilitating the improvement of the image recognition precision and reliability of an image recognition task. In addition, the method is high in processing speed, is small in size, and can be conveniently arranged on mobile equipment, such as a smartphone, a tablet PC, and other devices.

Description

Image shadow detection method and device

Technical field

The present invention relates to image processing field, relate more specifically to a kind of image shadow detection method and device.

Background technology

The moon for many image recognition tasks (such as image classification, Text region, target following etc.), in image Shadow is a serious disturbing factor, can affect the precision and stability (i.e. reliability) of image recognition.If it is possible to Judge before execution image recognition tasks to whether there is shade in image and predict the region at shade place it is possible to effectively drop The difficulty of low image recognition tasks, the precision and stability of lifting image recognition.However, still lack at present to detect exactly The mature technology of the shade in image or system.

Content of the invention

Propose the present invention in view of the problems referred to above.The invention provides a kind of image shadow detection method and device.

According to an aspect of the present invention, there is provided a kind of image shadow detection method.This image shadow detection method includes：Obtain Take altimetric image to be checked；And using full convolutional network, described altimetric image to be checked is processed, to obtain with regard to described to be detected The testing result of the positional information of the shadow region in image.

Exemplarily, described testing result includes shade probability graph, the pixel of each pixel in described shade probability graph Value represents the probability that there is shade in described altimetric image to be checked and this pixel coordinate identical pixel.

Exemplarily, described image shadow detection method also includes：Obtain training data, wherein, described training data bag Include sample graph image set and each sample image corresponding mask information respectively concentrated with described sample image, described mask information For indicating the position of the shadow region in corresponding sample image；And carry out neural metwork training using described training data To obtain described full convolutional network.

Exemplarily, described mask information is the bianry image equivalently-sized with corresponding sample image, described binary map As in be located at corresponding sample image shadow region within pixel coordinate identical pixel there is the first pixel value, And, in described bianry image be located at corresponding sample image shadow region outside pixel coordinate identical pixel There is the second pixel value.

Exemplarily, described acquisition training data includes：Obtain initial image set, wherein, described initial pictures are concentrated The sample image that initial pictures are concentrated with described sample image corresponds；Each initial graph that described initial pictures are concentrated Picture, generates the shadow region of predetermined number；Each initial pictures concentrated for described initial pictures, by this initial pictures and institute The shadow region stating predetermined number is superimposed, to obtain the sample image corresponding with this initial pictures；And for institute State each initial pictures of initial pictures concentration, superposition position in this initial pictures for the shadow region according to described predetermined number Put and obtain the mask information corresponding with the sample image corresponding to this initial pictures.

Exemplarily, the described shadow region generating predetermined number includes：Generate multiple images block；From the plurality of image The image block that number is equal to described predetermined number is randomly choosed in block；And by each image block in selected image block Pixel the pixel value value that is randomly set in the range of default shading value, to obtain the shadow region of described predetermined number.

Exemplarily, described using described training data carry out neural metwork training with obtain described full convolutional network it Before, described image shadow detection method also includes：The size scaling of the sample image that described sample image is concentrated is gauge Very little.

Exemplarily, the size scaling of the described sample image concentrating described sample image includes for standard size：Right Each sample image concentrated in described sample image, while the ratio of width to height keeping this sample image is constant, by this sample The greater in the height and width of image zooms to normal size.

Exemplarily, described acquisition training data includes：Receive described sample graph image set and concentrate with described sample image Each sample image respectively corresponding labeled data, wherein, described labeled data is included for indicating corresponding sample image In the profile of shadow region outline data；And according to corresponding with each sample image that described sample image is concentrated Outline data generates the mask information corresponding with this sample image.

Exemplarily, described using full convolutional network, described altimetric image to be checked is processed before, described image is cloudy Shadow detection method also includes：The size scaling of described altimetric image to be checked is standard size.

Exemplarily, the described size scaling by described altimetric image to be checked includes for standard size：Keeping described to be checked While the ratio of width to height of altimetric image is constant, the greater in the height and width of described altimetric image to be checked is zoomed to standard big Little.

Exemplarily, described using full convolutional network, described altimetric image to be checked is carried out process include：Will be standard-sized Described altimetric image to be checked inputs described full convolutional network, to obtain with regard to the shadow region in standard-sized described altimetric image to be checked The Primary Outcome of the positional information in domain；And described in the scaling according to described altimetric image to be checked and the acquisition of described Primary Outcome Testing result.

Exemplarily, described full convolutional network is configured to there is following network structure：Input layer, followed by two convolution Layer, connects a maximum pond layer, followed by two convolutional layers, connects a maximum pond layer, followed by three convolution Layer, connects a maximum pond layer, followed by three convolutional layers, connects a maximum pond layer, followed by three convolution Layer.

According to a further aspect of the invention, there is provided a kind of image shadow Detection device, including：Image collection module, is used for Obtain altimetric image to be checked；And processing module, for being processed to described altimetric image to be checked using full convolutional network, to obtain Testing result with regard to the positional information of the shadow region in described altimetric image to be checked.

Exemplarily, described image shadow Detection device also includes：Data acquisition module, for obtaining training data, its In, described training data includes sample graph image set and each sample image corresponding mask respectively concentrated with described sample image Information, described mask information is used for indicating the position of the shadow region in corresponding sample image；And training module, for profit Carry out neural metwork training with described training data to obtain described full convolutional network.

Exemplarily, described data acquisition module includes：Initial pictures acquisition submodule, for obtaining initial image set, Wherein, the sample image that the initial pictures that described initial pictures are concentrated are concentrated with described sample image corresponds；Shadow region Generate submodule, for each initial pictures concentrated for described initial pictures, generate the shadow region of predetermined number；Superposition Submodule, for each initial pictures concentrated for described initial pictures, by the moon of this initial pictures and described predetermined number Shadow zone domain is superimposed, to obtain the sample image corresponding with this initial pictures；And mask information obtains submodule, use In each initial pictures concentrated for described initial pictures, the shadow region according to described predetermined number is in this initial pictures Superposed positions obtain the mask information corresponding with the sample image corresponding to this initial pictures.

Exemplarily, described shadow region generates submodule and includes：Image block signal generating unit, for generating multiple images Block；Image block select unit, for randomly choosing the image block that number is equal to described predetermined number from the plurality of image block； And pixel value setup unit, for the pixel value of the pixel in each image block in selected image block is randomly provided For presetting the value in the range of shading value, to obtain the shadow region of described predetermined number.

Exemplarily, described image shadow Detection device also includes：First Zoom module, in described training module profit With described training data carry out neural metwork training to obtain described full convolutional network before, by described sample image concentrate sample The size scaling of this image is standard size.

Exemplarily, described first Zoom module includes：First scaling submodule, for concentrating for described sample image Each sample image, keep this sample image the ratio of width to height constant while, by the height and width of this sample image The greater zoom to normal size.

Exemplarily, described data acquisition module includes：Receiving submodule, for receive described sample graph image set and with institute State each sample image corresponding labeled data respectively of sample image concentration, wherein, described labeled data is included for indicating The outline data of the profile of the shadow region in corresponding sample image；And mask information generate submodule, for according to The outline data that each sample image of described sample image concentration is corresponding generates the mask letter corresponding with this sample image Breath.

Exemplarily, described image shadow Detection device also includes：Second Zoom module, in described processing module profit Before described altimetric image to be checked being processed with full convolutional network, the size scaling of described altimetric image to be checked is gauge Very little.

Exemplarily, described second Zoom module includes：Second scaling submodule, for keeping described altimetric image to be checked The ratio of width to height constant while, the greater in the height and width of described altimetric image to be checked is zoomed to normal size.

Exemplarily, described processing module includes：Input submodule, for will be defeated for standard-sized described altimetric image to be checked Enter described full convolutional network, first with regard to the positional information of the shadow region in standard-sized described altimetric image to be checked with acquisition Level result；And testing result obtains submodule, for the scaling according to described altimetric image to be checked and described Primary Outcome Obtain described testing result.

Image shadow detection method according to embodiments of the present invention and device, can be effective, high-precision using full convolutional network Shadow region in degree ground detection image, thus be favorably improved precision and the reliability of the image recognition in image recognition tasks Property.Additionally, the method also has the characteristics that processing speed is fast, model small volume, therefore can easily be deployed to such as intelligent On the mobile devices such as mobile phone, panel computer.

Brief description

By combining accompanying drawing, the embodiment of the present invention is described in more detail, the above-mentioned and other purpose of the present invention, Feature and advantage will be apparent from.Accompanying drawing is used for providing the embodiment of the present invention is further understood, and constitutes explanation A part for book, is used for explaining the present invention together with the embodiment of the present invention, is not construed as limiting the invention.In the accompanying drawings, Identical reference number typically represents same parts or step.

Fig. 1 illustrates the exemplary electronic device for realizing image shadow detection method according to embodiments of the present invention and device Schematic block diagram；

Fig. 2 illustrates the indicative flowchart of image shadow detection method according to an embodiment of the invention；

Fig. 3 illustrates altimetric image to be checked according to an embodiment of the invention and the schematic diagram of corresponding shade probability graph；

Fig. 4 illustrates the schematic block diagram of the training step of full convolutional network according to an embodiment of the invention；

The flow chart that Fig. 5 illustrates the step obtaining training data according to an embodiment of the invention；

Fig. 6 shows the schematic block diagram of image shadow Detection device according to an embodiment of the invention；And

Fig. 7 illustrates the schematic block diagram of image shadow Detection system according to an embodiment of the invention.

Specific embodiment

So that the object, technical solutions and advantages of the present invention become apparent from, describe root below with reference to accompanying drawings in detail Example embodiment according to the present invention.Obviously, described embodiment is only a part of embodiment of the present invention, rather than this Bright whole embodiments are not it should be appreciated that the present invention is limited by example embodiment described herein.Described in the present invention The embodiment of the present invention, the obtained all other embodiment in the case of not paying creative work of those skilled in the art All should fall under the scope of the present invention.

In order to solve problem as described above, the embodiment of the present invention propose a kind of based in full convolutional network detection image The method and apparatus of shade.

First, to describe with reference to Fig. 1 for realizing image shadow detection method according to embodiments of the present invention and device Exemplary electronic device 100.

As shown in figure 1, electronic equipment 100 includes one or more processors 102, one or more storage device 104, defeated Enter device 106, output device 108 and image collecting device 110, these assemblies pass through bus system 112 and/or other forms Bindiny mechanism's (not shown) interconnection.It should be noted that the assembly of electronic equipment 100 shown in Fig. 1 and structure are exemplary, and Nonrestrictive, as needed, described electronic equipment can also have other assemblies and structure.

Described processor 102 can be CPU (CPU) or have data-handling capacity and/or instruction execution The processing unit of the other forms of ability, and the other assemblies in described electronic equipment 100 can be controlled desired to execute Function.

Described storage device 104 can include one or more computer programs, and described computer program can To include various forms of computer-readable recording mediums, such as volatile memory and/or nonvolatile memory.Described easy The property lost memory for example can include random access memory (RAM) and/or cache memory (cache) etc..Described non- Volatile memory for example can include read-only storage (ROM), hard disk, flash memory etc..In described computer-readable recording medium On can store one or more computer program instructions, processor 102 can run described program instruction, to realize hereafter institute The client functionality (realized by processor) in the embodiment of the present invention stated and/or other desired function.In described meter Various application programs and various data can also be stored in calculation machine readable storage medium storing program for executing, such as described application program using and/or Various data producing etc..

Described input unit 106 can be the device for input instruction for the user, and can include keyboard, mouse, wheat Gram one or more of wind and touch-screen etc..

Described output device 108 can export various information (such as image and/or sound) to outside (such as user), and And one or more of display, loudspeaker etc. can be included.

Described image harvester 110 can gather the altimetric image to be checked for shadow Detection, and by the figure being gathered As being stored in described storage device 104 so that other assemblies use.Image collecting device 110 can be camera.Should manage Solution, image collecting device 110 is only example, and electronic equipment 100 can not include image collecting device 110.In this case, Other image acquisition device altimetric image to be checked can be utilized, and the image of collection is sent to electronic equipment 100, or electricity Sub- equipment 100 can be downloaded via network or directly obtain to be checked from local storage (for example above-mentioned storage device 104) Altimetric image.

Exemplarily, the exemplary electron for realizing image shadow detection method according to embodiments of the present invention and device sets Standby can have realization on the equipment that data calculating is with disposal ability various, and for example, it can be in such as smart mobile phone, flat board Realize on the mobile device of computer etc., personal computer or remote server.

According to one aspect of the invention, provide a kind of image shadow detection method.Fig. 2 illustrates according to one enforcement of the present invention The schematic block diagram of the image shadow detection method 200 of example.As shown in Fig. 2 image shadow detection method 200 includes following step Suddenly.

In step S210, obtain altimetric image to be checked.

Altimetric image to be checked can be any image needing detection shade, for example, be used for the image of image recognition.To be detected Image can be imaged the original image that arrives of first-class image acquisition device or download or locally stored via network Original image or original image is carried out pre-processing the image obtaining afterwards.

In step S220, treat detection image using full convolutional network and processed, to obtain with regard in altimetric image to be checked The positional information of shadow region testing result.

Image can represent in the color spaces such as RGB, HSV.For the image that there is shade, its shadow region is in HSV The luminance component in HSV space is low generally than shadeless region for luminance component in space, therefore, exemplarily, can To determine the scope of shadow region according to the brightness value of pixel each in image.That is, " shade " and " shadow region " is permissible Parameter based on such as pixel brightness value being divided and to be defined, for example, it is possible to by whole image be divided into shadow region and Non-hatched area.

The full convolutional network that altimetric image to be checked input can be trained.Full convolutional network can be with semantic segmentation model M Represent, model M includes network structure and its parameter of full convolutional network.Through the process of full convolutional network, in full convolutional network Output end can obtain the testing result of the positional information with regard to the shadow region in altimetric image to be checked.

According to one embodiment of the invention, testing result can include shade probability graph P, in described shade probability graph P The pixel value of each pixel represents the probability that there is shade in altimetric image to be checked and this pixel coordinate identical pixel.

Shade probability graph P is identical with picture size to be detected, and it is mutually right that the coordinate identical pixel of the two can be considered as The pixel answered.In one example, shade probability graph P is gray level image, and the pixel value of its pixel is the gray value of its pixel. In the case of representing shadow region using inclined white pixel and non-hatched area represented using inclined black picture element, shade probability graph The pixel value of certain pixel in P is bigger, and the probability that there is shade at the respective pixel representing altimetric image to be checked is bigger.Fig. 3 illustrates Altimetric image to be checked according to an embodiment of the invention and the schematic diagram of corresponding shade probability graph.With reference to Fig. 3, shown in left side It is original altimetric image to be checked, it is the image for identity card collection, right side is illustrated that to be processed using full convolutional network and treats The testing result of output, i.e. described shade probability graph after detection image.As described in Figure 3, on a left side for the altimetric image to be checked in left side There is one piece of shadow region, the corresponding position in the shade probability graph on right side is rendered as obvious white, remaining at inferior horn Position corresponding with non-hatched area is then rendered as obvious black, the correspondence position of shadow region and the correspondence of non-hatched area Boundary between position is apparent from.In the embodiment shown in fig. 3, the pixel value of certain pixel in shade probability graph P is bigger, generation The probability that there is shade at the respective pixel of table altimetric image to be checked is bigger.Meanwhile, by the shade probability graph P shown in Fig. 3, permissible Clear and definite ground determines the position of the shadow region in altimetric image to be checked.

It is, of course, understood that shadow region being represented using inclined black picture element and representing non-using inclined white pixel In the case of shadow region, the pixel value of certain pixel in shade probability graph P is less, represents at the respective pixel of altimetric image to be checked The probability that there is shade is bigger, and this is contrary with a kind of upper situation.In this case, the black of the shade probability graph on the right side of Fig. 3 The pixel color of part and white portion will be exchanged.

Image shadow detection method according to embodiments of the present invention, using the full convolutional network training, can directly examine Survey shade whether there is, and can accurately be partitioned into the region at shade place, therefore the method has high precision, adaptability Strong feature, can be greatly enhanced the precision of image recognition in associated picture identification mission, reliability.Additionally, according to this The image shadow detection method of inventive embodiments also has the characteristics that processing speed is fast, model small volume, therefore can be easily It is deployed on the mobile devices such as smart mobile phone, panel computer.

Exemplarily, image shadow detection method according to embodiments of the present invention can have memory and processor Realize in unit or system.

Image shadow detection method according to embodiments of the present invention can be deployed at IMAQ end, for example, it is possible to portion Administration is in mobile phone end.Alternatively, image shadow detection method according to embodiments of the present invention can also be deployed in clothes with being distributed At business device end (or high in the clouds) and client.For example, it is possible to gather altimetric image to be checked in client, client is to be checked by collect Altimetric image sends server end (or high in the clouds) to, carries out image shadow Detection by server end (or high in the clouds).

According to embodiments of the present invention, image shadow detection method 200 can also include the training step of neutral net.Fig. 4 The schematic block diagram of the training step S400 of neutral net according to an embodiment of the invention is shown.

As shown in figure 4, the training step S400 of neutral net may comprise steps of.

In step S410, obtain training data, wherein, described training data include sample graph image set and with sample graph image set In each sample image respectively corresponding mask information, described mask information is used for indicating the shade in corresponding sample image The position in region.

Sample image can be any image suitably comprising shade.Sample image can be camera collect former Beginning image or download via network or locally stored original image or after original image is pre-processed The image obtaining.

Sample graph image set can include any number of sample image, including but not limited to only includes a sample image Situation.Exemplarily, a large amount of (such as more than 5000) sample images can be collected in advance, and will be defeated for the sample image collected Enter electronic equipment 100, neural metwork training is carried out using these sample images by electronic equipment 100.

Each sample image has corresponding mask information.In one example, sample image can be acquired original Comprise the natural image of shade, in such a case, it is possible to what the shade being marked out using manual type in sample image was located Region, to obtain labeled data, mask information can be generated based on this labeled data.In another example, sample image Can be the composograph of the inclusion shadow region being synthesized using the image not comprising or containing substantially no shade, in this situation Under, the mask information corresponding with each sample image can be generated according to synthesis situation.

Mask information can have any suitable form, and it is mainly used in indicating the shadow region in corresponding sample image The position in domain.In one example, mask information can be the wheel of the profile with regard to the shadow region in corresponding sample image Wide data.Outline data can include the coordinate of the point on the profile of shadow region.Mark the profile of shadow region it is possible to Know the position of shadow region.In another example, mask information can be the two-value equivalently-sized with corresponding sample image Image, this example will be described in more detail below.The mask information of above-mentioned form is only exemplary rather than limiting, and the invention is not restricted to This.

In step S420, carry out neural metwork training to obtain described full convolutional network using training data.

Using each sample image and mask information (such as bianry image) corresponding thereto as input, carry out nerve net The training of network, to obtain described full convolutional network.Full convolutional network has initial parameter, and these initial parameters can be experience Value, the initial parameter of convolutional network is thus continually updated entirely in the training process, final needed for can obtaining when training completes Parameter.In the step s 420, for example can be passed through reversely with being trained to full convolutional network using conventional neural metwork training mode The methods such as propagation algorithm are being trained.Can be used for the inspection of image shade in the full convolutional network obtaining final after training Survey.

By above training method, acquisition can be trained to have the full convolutional network compared with high measurement accuracy.

According to embodiments of the present invention, mask information is the bianry image equivalently-sized with corresponding sample image, described two In value image be located at corresponding sample image shadow region within pixel coordinate identical pixel there is the first pixel Value, and, in described bianry image be located at corresponding sample image shadow region outside pixel coordinate identical picture Element has the second pixel value.

Mask information can be obtained by carrying out binary conversion treatment to sample image.Image binaryzation is exactly by image The gray value of pixel be set to 0 or 255, that is, make whole image present obvious black and white effect.In an example In, can be by the gray value of the pixel within the shadow region in sample image be set to 255, by the gray scale of rest of pixels Value is set to 0 to obtain bianry image.So, in bianry image, the place that there is shade is rendered as white, and remainder is in It is now black, the boundary between shadow region and non-hatched area is very clear, thus shadow region is split.According to this The bianry image that example obtains is on assuming effect similar to the image right shown in Fig. 3.In another example, can pass through will The gray value of the pixel within shadow region in sample image is set to 0, and the gray value of rest of pixels is set to 255 to obtain Obtain bianry image.So, in bianry image, the place that there is shade is rendered as black, and remainder is rendered as white, equally Shadow region can be split.

As can be seen here, bianry image is equivalently-sized with sample image, pixel correspond it is believed that bianry image be by Sample image is transformed, and the difference of the two is that bianry image mainly highlights shadow region and to included in image Remaining information is not concerned with.

Due to only comprising two kinds of gray value data in bianry image, therefore image is very simple, and data volume is smaller, and The profile of target interested can be highlighted.In addition, the amount of calculation required when processing to bianry image is also smaller. Therefore, the position to record the shadow region in sample image using bianry image is the simply efficient mode of one kind, after being easy to Continue the training of full convolutional network.

As described above, sample image can obtain by way of image synthesizes, and is described with reference to Fig. 5.Fig. 5 The flow chart illustrating the step (step S410) obtaining training data according to an embodiment of the invention.According to the present embodiment, Step S410 may comprise steps of.

In step S412, obtain initial image set, wherein, initial pictures and sample image concentration that initial pictures are concentrated Sample image corresponds.

Initial pictures can be any suitable natural image, and it can comprise as far as possible few shade, compare preferably Initial pictures do not comprise shade completely.The shade comprising in initial pictures is few, so that the full convolutional network being trained Accuracy of detection is higher.Initial pictures can be to image the original image that arrives of first-class image acquisition device or via network Download or locally stored original image or original image is carried out pre-processing the image obtaining afterwards.

A large amount of (such as more than 5000) initial pictures can be collected, these initial pictures can be designated as set S={ I₁, I₂,...,I_N, i.e. initial image set, wherein N represent the number of initial pictures.

In step S414, each initial pictures concentrated for initial pictures, generate the shadow region of predetermined number.

For each initial pictures I in set S_k, k=1 ..., N, H shadow region can be generated, H is herein Described predetermined number, it is configurable parameter.The number of the shadow region generated in step S414 can set as needed Fixed, the present invention is not limited to this.In one example, the span of predetermined number can be [1,5].It should be appreciated that It is that the span of predetermined number can set according to the actual requirements.The region that generally there is shade in piece image may Will not be too many, so the number of the shadow region generated in step S414 need not be too big.Initial pictures concentration is appointed For the different initial pictures of meaning two, the number (i.e. predetermined number) of the shadow region being generated can identical it is also possible to different.

In one example, step S414 can include：Generate multiple images block；Select at random from the plurality of image block Select the image block that number is equal to described predetermined number；And the picture by the pixel in each image block in selected image block Plain value is randomly set to the value in the range of default shading value, to obtain the shadow region of predetermined number.

The number of the image block being generated can be consistent with the number of required shadow region, or the image block being generated Number can be more than the number of required shadow region.Then number can be selected from the image block being generated to be equal to predetermined The image block of number, so selected image block is corresponded with the shadow region of predetermined number.Number is equal to predetermined number Image block can be obtained by choosing in any way on image (can be any blank image).For example, it is possible to Random generation straight line in blank image, this straight line divides an image into two parts, then can randomly choose therein one Part is equal to the image block of predetermined number as number, and in this case, predetermined number is 1.If straight line is marked off Two parts all as number be equal to predetermined number image block, then predetermined number be 2.According to another example, can be in blank The figure of arbitrary shape generated on image, for example circular, square, triangle etc., the image block that the profile of this figure is surrounded For required image block.Desired predetermined number is how many just can generate how many such figures.For example, it is assumed that it is pre- Fixed number mesh is 5, then can generate 5 figures, the shape of any two of which different graphic and/or size can be identical, also may be used With difference.

Default shading value scope can be any suitable scope, and the present invention is not limited to this.For example, default shade Scope can be [- 255, -128], and the value in the range of this refers to gray value.The pixel value of each pixel in image block is permissible It is randomly provided, that is, the pixel value of different pixels may be different.It should be understood, however, that can be by same image All pixels in block are set to identical pixel value, furthermore it is also possible to the pixel in any two different images block is arranged For identical pixel value.

By with upper type, the shadow region of predetermined number can be automatically generated.

In step S416, each initial pictures concentrated for initial pictures, by the moon of this initial pictures and predetermined number Shadow zone domain is superimposed, to obtain the sample image corresponding with this initial pictures.

Shadow region is superimposed with initial pictures, sample image can be synthesized.The superposition position of each shadow region Put and can arbitrarily set.By the shadow region generating and initial pictures I_kIt is overlapped, obtain image (the i.e. sample graph synthesizing Picture) I '_k, the image collection of multiple synthesis together, constitute sample graph image set.

In step S418, each initial pictures concentrated for initial pictures, the shadow region according to predetermined number is at this Superposed positions in initial pictures obtain the mask information corresponding with the sample image corresponding to this initial pictures.

Determining that shadow region (can be in shadow region and initial graph in the case of the superposed positions in initial pictures As before superposition, afterwards or meanwhile), mask information as herein described can be determined according to the superposed positions of shadow region. For example, in the case that mask information is bianry image, can by sample image, in the superposed positions of shadow region The gray value of pixel is set to 255, and in sample image, pixel outside the superposed positions of shadow region gray value is arranged For 0, thus obtaining bianry image L_k.By the bianry image L being obtained with upper type_kIt is one and sample image I '_kEquivalently-sized Image, for I '_kThere is the region of shade, L_kThe gray value of the pixel of middle correspondence position is 255, the pixel of remaining position Gray value is 0.

One set T={ (I ' can be constructed by above step₁,L₁),(I′₂,L₂),...,(I′_N,L_N), described collection Close each sample image corresponding mask information respectively including sample graph image set and concentrating with described sample image, this set T It is training data as herein described.

To obtain required training data and can remove loaded down with trivial details artificially collecting by the way of being automatically synthesized sample image from With the process of flag data, time cost and cost of labor can be reduced.Further, since shadow region is to calculate to generate, because This its location comparison is controlled, and the mask information degree of accuracy for indicating the position of shadow region also can be higher, so that instruction The accuracy of detection of the full convolutional network practised is higher.

According to embodiments of the present invention, before step S420, image shadow detection method 200 can also include：By sample The size scaling of the sample image in image set is standard size.

Full convolutional network can process the image of sizes, and therefore, it can will be directly defeated for the sample image of original size Enter full convolutional network to be trained.It is, of course, understood that can also before sample image is inputted full convolutional network, The size scaling of sample image is standard size, sample image is normalized.Image is too little to be not easy to identify shade Region, image too macrooperation amount is larger, therefore sample image can be normalized to suitable size and input full convolutional network again It is trained.

According to embodiments of the present invention, the size scaling of sample image sample image concentrated can wrap for standard size Include：Each sample image concentrated for sample image, while the ratio of width to height keeping this sample image is constant, by this sample The greater in the height and width of image zooms to normal size.

For each sample image, the height of sample image can be compared with width, in the two relatively Big person, scales it normal size, the ratio of width to height is constant simultaneously.Described normal size can set as needed, and for example it can Think 160 pixels, 192 pixels, 224 pixels, 256 pixels etc..

It should be noted that above-mentioned mask information be bianry image embodiment in, bianry image will and sample image Size be consistent.Therefore, if sample image is scaled, bianry image can be scaled accordingly, or Person after sample image zooms in and out, then can generate bianry image based on the sample image after scaling.

According to one embodiment of the invention, shadow region can be obtained by way of mark.In this case, step S410 can include：Receive sample graph image set and each sample image corresponding labeled data respectively concentrated with sample image, Wherein, described labeled data includes the outline data of the profile for indicating the shadow region in corresponding sample image；And The mask corresponding with this sample image is generated according to the outline data corresponding with each sample image that sample image is concentrated Information.

As described above, training data can be obtained by way of artificially collecting and marking.In this case, sample Image can be the natural image comprising shade.After collecting great amount of samples image, can be by mark personnel by each sample Shadow region on this image marks out, for example, typically marks the point on the profile of shadow region, thus can obtain mark Note data.Electronic equipment 100, after receiving sample image and corresponding labeled data, can determine cloudy according to labeled data Position in sample image for the shadow zone domain, such that it is able to generate mask information.For example, sample image and corresponding mark are being received After note data, can be according to labeled data by sample image, contoured interior in shadow region pixel gray value It is set to 255, be set to 0 by sample image, in the gray value of the pixel of the profile exterior of shadow region, to obtain two-value Image.

According to one embodiment of the invention, described full convolutional network is configured to there is following network structure：Input layer, connects down To be two convolutional layers, to connect a maximum pond layer, followed by two convolutional layers, to connect a maximum pond layer, connect down To be three convolutional layers, to connect a maximum pond layer, followed by three convolutional layers, to connect a maximum pond layer, connect down Three convolutional layers.

Full convolutional network is a kind of deep neural network structure, mainly by convolutional layer (convolutional layer), pond Change layer (pooling layer) and up-sampling layer (up-sampling layer) composition.In the present embodiment, using of equal value Convolutional layer replaces the full articulamentum (fully-connected layer) in the full convolutional network of tradition, and this can be to a certain degree Upper simplified operation, reduces data amount of calculation.Certainly, full convolutional network according to embodiments of the present invention can not also adopt equivalence Convolutional layer is replaced full articulamentum but is still realized using conventional full articulamentum.

The input layer of full convolutional network receives training data or altimetric image to be checked；Followed by two convolutional layers, exemplary Ground, the number of the wave filter of each convolutional layer in this two convolutional layers is 6, and wave filter size is 3x3；Subsequently, connect one Maximum pond layer (max-pooling layer)；Followed by two convolutional layers, exemplarily, every in this two convolutional layers The number of the wave filter of individual convolutional layer is 12, and wave filter size is 3x3；Subsequently, connect a maximum pond layer；Followed by three Individual convolutional layer, exemplarily, the number of the wave filter of each convolutional layer in these three convolutional layers is 16, and wave filter size is 3x3；Subsequently, connect a maximum pond layer；Followed by three convolutional layers, the filter of each convolutional layer in these three convolutional layers The number of ripple device is 24, and wave filter size is 3x3；Subsequently, connect a maximum pond layer；Followed by three convolutional layers, this The number of the wave filter of each convolutional layer in three convolutional layers is 32, and wave filter size is 3x3.

Sample set T can be sent into the full convolutional network designing to be trained, obtain above-mentioned model M.Training Cheng Zhong, every time a sample image and corresponding mask information is input in model M, and initial learning rate is 0.00000001, Often take turns iteration through 10000, learning rate is reduced to original 1/10.After iteration 100000 wheel, training process terminates, it is possible to obtain The full convolutional network training.Subsequently, it is possible to train altimetric image to be checked input during shade in needing detection image Full convolutional network processed, export shade probability graph mentioned above.

It should be appreciated that the network structure of above-mentioned full convolutional network is merely illustrative and unrestricted, full convolutional network can have Other any suitable network structures, the layer such as convolutional layer therein, pond layer can have other suitable numbers, and every layer In wave filter can also have other suitable numbers and size.Additionally, above-mentioned maximum pond layer can adopt average pond The pond layer of the other forms such as layer replaces.

According to embodiments of the present invention, before step S220, image shadow detection method 200 can also include：Will be to be checked The size scaling of altimetric image is standard size.

As described above, full convolutional network can process the image of sizes, therefore, it can original size is to be checked Altimetric image directly inputs full convolutional network and is processed.Certainly, with sample image is zoomed in and out similarly, using full convolution Zoom in and out it is also possible to treat detection image during network processes altimetric image to be checked.Similarly, due to the too little inconvenience of image In identification shadow region, image too macrooperation amount is larger, can be therefore that suitable size is defeated again by image normalization to be detected Enter full convolutional network to be processed.It should be noted that the scaling of altimetric image to be checked can be with the contracting of arbitrary sample image Ratio of putting is identical or different.It is to be appreciated that for the different sample images of sample image concentration, its scaling also may be used With identical or different.

According to embodiments of the present invention, the described size scaling by altimetric image to be checked can include for standard size：Keeping While the ratio of width to height of altimetric image to be checked is constant, the greater in the height and width of altimetric image to be checked is zoomed to standard big Little.

The height of altimetric image to be checked can be compared with width, for the greater in the two, scale it standard Size, the ratio of width to height is constant simultaneously.Described normal size can set as needed, for example its can for 160 pixels, 192 pixels, 224 pixels, 256 pixels etc..

According to embodiments of the present invention, step S220 can include：Standard-sized altimetric image to be checked is inputted full convolution net Network, to obtain the Primary Outcome of the positional information with regard to the shadow region in standard-sized altimetric image to be checked；And according to treating The scaling of detection image and Primary Outcome obtain testing result.

If treating detection image before processing altimetric image to be checked using full convolutional network to zoom in and out, due to be detected The size of image creates change, and therefore, the result (i.e. Primary Outcome) of full convolutional network output is corresponding to be waiting after scaling Detection image and nonprimitive altimetric image to be checked, therefore can the scaling according to altimetric image to be checked and Primary Outcome further Obtain and the original corresponding testing result of altimetric image to be checked.For example, it is assumed that full convolutional network output is shade probability graph, and And assume that altimetric image to be checked is enlarged into original 2 times, then the shade probability graph that full convolutional network directly can be exported is carried out The scaling contrary with altimetric image to be checked, that is, be reduced into original 2 times, thus obtaining final testing result, its be pixel with former The pixel one-to-one shade probability graph of the altimetric image to be checked beginning.

According to a further aspect of the invention, provide a kind of image shadow Detection device.Fig. 6 shows according to one reality of the present invention Apply the schematic block diagram of the image shadow Detection device 600 of example.

As shown in fig. 6, image shadow Detection device 600 according to embodiments of the present invention includes image collection module 610 He Processing module 620.Described modules can execute each of the image shadow detection method above in conjunction with Fig. 2-5 description respectively Step/function.Hereinafter only the major function of each module of this image shadow Detection device 600 is described, and more than omitting The detail content having been noted above.

Image collection module 610 is used for obtaining altimetric image to be checked.Image collection module 610 can electronics as shown in Figure 1 In processor 102 Running storage device 104 in equipment, the programmed instruction of storage is realizing.

Processing module 620 is processed for treating detection image using full convolutional network, to obtain with regard to mapping to be checked The testing result of the positional information of shadow region in picture.Processing module 620 can process in electronic equipment as shown in Figure 1 In device 102 Running storage device 104, the programmed instruction of storage is realizing.

According to embodiments of the present invention, described testing result includes shade probability graph, each picture in described shade probability graph The pixel value of element represents the probability that there is shade in described altimetric image to be checked and this pixel coordinate identical pixel.

According to embodiments of the present invention, described image shadow Detection device 600 also includes：Data acquisition module, for obtaining Training data, wherein, described training data includes sample graph image set and each sample image with described sample image concentration divides Not corresponding mask information, described mask information is used for indicating the position of the shadow region in corresponding sample image；And instruction Practice module, for carrying out neural metwork training to obtain described full convolutional network using described training data.

According to embodiments of the present invention, described mask information is the bianry image equivalently-sized with corresponding sample image, institute State in bianry image be located at corresponding sample image shadow region within pixel coordinate identical pixel have first Identical with the pixel coordinate outside the shadow region being located at corresponding sample image in pixel value, and, described bianry image Pixel there is the second pixel value.

According to embodiments of the present invention, described data acquisition module 610 includes：Initial pictures acquisition submodule, for obtaining Initial image set, wherein, a pair of the sample image 1 that initial pictures that described initial pictures are concentrated are concentrated with described sample image Should；Shadow region generates submodule, for each initial pictures concentrated for described initial pictures, generates the moon of predetermined number Shadow zone domain；Superposition submodule, for each initial pictures concentrated for described initial pictures, this initial pictures are pre- with described Fixed number purpose shadow region is superimposed, to obtain the sample image corresponding with this initial pictures；And mask information obtains Obtain submodule, for each initial pictures concentrated for described initial pictures, existed according to the shadow region of described predetermined number Superposed positions in this initial pictures obtain the mask information corresponding with the sample image corresponding to this initial pictures.

According to embodiments of the present invention, described shadow region generates submodule and includes：Image block signal generating unit is many for generating Individual image block；Image block select unit, is equal to described predetermined number for randomly choosing number from the plurality of image block Image block；And pixel value setup unit, for by the pixel value of the pixel in each image block in selected image block It is randomly set to the value in the range of default shading value, to obtain the shadow region of described predetermined number.

According to embodiments of the present invention, described image shadow Detection device 600 also includes：First Zoom module, in institute State training module using described training data carry out neural metwork training to obtain described full convolutional network before, by described sample The size scaling of the sample image in image set is standard size.

According to embodiments of the present invention, described first Zoom module includes：First scaling submodule, for for described sample Each sample image in image set, while the ratio of width to height keeping this sample image is constant, by the height of this sample image Zoom to normal size with the greater in width.

According to embodiments of the present invention, described data acquisition module 610 includes：Receiving submodule, for receiving described sample Image set and each sample image corresponding labeled data respectively concentrated with described sample image, wherein, described labeled data Outline data including the profile for indicating the shadow region in corresponding sample image；And mask information generates submodule Block, for generating and this sample image phase according to the outline data corresponding with each sample image that described sample image is concentrated Corresponding mask information.

According to embodiments of the present invention, described image shadow Detection device 600 also includes：Second Zoom module, in institute State before processing module processed to described altimetric image to be checked using full convolutional network, by the size contracting of described altimetric image to be checked Put as standard size.

According to embodiments of the present invention, described second Zoom module includes：Second scaling submodule, for treating described in holding While the ratio of width to height of detection image is constant, the greater in the height and width of described altimetric image to be checked is zoomed to standard big Little.

According to embodiments of the present invention, described processing module 620 includes：Input submodule, for will be standard-sized described Altimetric image to be checked inputs described full convolutional network, to obtain with regard to the shadow region in standard-sized described altimetric image to be checked The Primary Outcome of positional information；And testing result obtains submodule, for the scaling according to described altimetric image to be checked and Described Primary Outcome obtains described testing result.

According to embodiments of the present invention, described full convolutional network is configured to there is following network structure：Input layer, followed by Two convolutional layers, connect a maximum pond layer, followed by two convolutional layers, connect a maximum pond layer, followed by Three convolutional layers, connect a maximum pond layer, followed by three convolutional layers, connect a maximum pond layer, followed by Three convolutional layers.

Those of ordinary skill in the art are it is to be appreciated that combine the list of each example of the embodiments described herein description Unit and algorithm steps, being capable of being implemented in combination in electronic hardware or computer software and electronic hardware.These functions are actually To be executed with hardware or software mode, the application-specific depending on technical scheme and design constraint.Professional and technical personnel Each specific application can be used different methods to realize described function, but this realization is it is not considered that exceed The scope of the present invention.

Fig. 7 shows the schematic block diagram of image shadow Detection system 700 according to an embodiment of the invention.Image is cloudy Shadow detecting system 700 includes image collecting device 710, storage device 720 and processor 730.

Image collecting device 710 is used for altimetric image to be checked.Image collecting device 710 is optional, image shadow Detection system System 700 can not include image collecting device 710.

Described storage device 720 stores corresponding in image shadow detection method according to embodiments of the present invention for realizing The program code of step.

Described processor 730 is used for running the program code of storage in described storage device 720, to execute according to the present invention The corresponding steps of the image shadow detection method of embodiment, and for realizing image shadow Detection according to embodiments of the present invention Image collection module 610 in device 600 and processing module 620.

In one embodiment, described program code makes described image shadow Detection system when being run by described processor 730 System 700 execution following steps：Obtain altimetric image to be checked；And using full convolutional network, described altimetric image to be checked is processed, To obtain the testing result of the positional information with regard to the shadow region in described altimetric image to be checked.

In one embodiment, described testing result includes shade probability graph, each pixel in described shade probability graph Pixel value represent the probability that there is shade in described altimetric image to be checked and this pixel coordinate identical pixel.

In one embodiment, described program code also makes described image shadow Detection when being run by described processor 730 System 700 executes：Obtain training data, wherein, described training data is included sample graph image set and concentrated with described sample image Each sample image respectively corresponding mask information, described mask information is used for indicating the shadow region in corresponding sample image The position in domain；And carry out neural metwork training to obtain described full convolutional network using described training data.

In one embodiment, described mask information is the bianry image equivalently-sized with corresponding sample image, described In bianry image be located at corresponding sample image shadow region within pixel coordinate identical pixel there is the first picture Element value, and, in described bianry image be located at corresponding sample image shadow region outside pixel coordinate identical Pixel has the second pixel value.

In one embodiment, described program code makes described image shadow Detection system when being run by described processor 730 The step of the acquisition training data performed by system 700 includes：Obtain initial image set, wherein, it is first that described initial pictures are concentrated The sample image that beginning image is concentrated with described sample image corresponds；Each initial graph that described initial pictures are concentrated Picture, generates the shadow region of predetermined number；Each initial pictures concentrated for described initial pictures, by this initial pictures and institute The shadow region stating predetermined number is superimposed, to obtain the sample image corresponding with this initial pictures；And for institute State each initial pictures of initial pictures concentration, superposition position in this initial pictures for the shadow region according to described predetermined number Put and obtain the mask information corresponding with the sample image corresponding to this initial pictures.

In one embodiment, described program code makes described image shadow Detection system when being run by described processor 730 The step of the shadow region of generation predetermined number performed by system 700 includes：Generate multiple images block；From the plurality of image block Middle random selection number is equal to the image block of described predetermined number；And by each image block in selected image block The pixel value of pixel is randomly set to the value in the range of default shading value, to obtain the shadow region of described predetermined number.

In one embodiment, make described image shadow Detection when described program code is run by described processor 730 Performed by system 700 using described training data carry out neural metwork training with obtain described full convolutional network step it Before, described program code also makes described image shadow Detection system 700 execute when being run by described processor 730：By described sample The size scaling of the sample image in this image set is standard size.

In one embodiment, described program code makes described image shadow Detection system when being run by described processor 730 The size scaling of the sample image concentrating described sample image performed by system 700 includes for standard-sized step：For Each sample image that described sample image is concentrated, while the ratio of width to height keeping this sample image is constant, by this sample graph The greater in the height and width of picture zooms to normal size.

In one embodiment, described program code makes described image shadow Detection system when being run by described processor 730 The step of the acquisition training data performed by system 700 includes：Receive described sample graph image set and with described sample image concentrate Each sample image corresponding labeled data respectively, wherein, described labeled data is included for indicating in corresponding sample image The profile of shadow region outline data；And according to the wheel corresponding with each sample image that described sample image is concentrated The wide data genaration mask information corresponding with this sample image.

In one embodiment, make described image shadow Detection when described program code is run by described processor 730 Before the step described altimetric image to be checked being processed using full convolutional network performed by system 700, described program code Described image shadow Detection system 700 is also made to execute when being run by described processor 730：Size contracting by described altimetric image to be checked Put as standard size.

In one embodiment, described program code makes described image shadow Detection system when being run by described processor 730 The size scaling by described altimetric image to be checked performed by system 700 includes for standard-sized step：Keeping described to be detected While the ratio of width to height of image is constant, the greater in the height and width of described altimetric image to be checked is zoomed to normal size.

In one embodiment, described program code makes described image shadow Detection system when being run by described processor 730 System 700 performed by using full convolutional network described altimetric image to be checked is carried out process include：Will be standard-sized described to be checked Altimetric image inputs described full convolutional network, to obtain the position with regard to the shadow region in standard-sized described altimetric image to be checked The Primary Outcome of information；And the scaling according to described altimetric image to be checked and described Primary Outcome obtain described detection knot Really.

In one embodiment, described full convolutional network is configured to there is following network structure：Input layer, followed by two Individual convolutional layer, connects a maximum pond layer, followed by two convolutional layers, connects a maximum pond layer, followed by three Individual convolutional layer, connects a maximum pond layer, followed by three convolutional layers, connects a maximum pond layer, followed by three Individual convolutional layer.

Additionally, according to embodiments of the present invention, additionally providing a kind of storage medium, storing program on said storage Instruction, when described program instruction is run by computer or processor for executing the image shadow Detection side of the embodiment of the present invention The corresponding steps of method, and for realizing the corresponding module in image shadow Detection device according to embodiments of the present invention.Described Storage medium for example can include the storage card of the smart phone, memory unit of panel computer, the hard disk of personal computer, read-only Memory (ROM), Erasable Programmable Read Only Memory EPROM (EPROM), portable compact disc read-only storage (CD-ROM), USB Memory or any combination of above-mentioned storage medium.

In one embodiment, described computer program instructions when being run by computer or processor so that calculating Each functional module of image shadow Detection device according to embodiments of the present invention realized by machine or processor, and/or permissible Execute image shadow detection method according to embodiments of the present invention.

In one embodiment, described computer program instructions make below described computer execution when being run by computer Step：Obtain altimetric image to be checked；And using full convolutional network, described altimetric image to be checked is processed, to obtain with regard to described The testing result of the positional information of the shadow region in altimetric image to be checked.

In one embodiment, described computer program instructions also make described computer execute when being run by computer： Obtain training data, wherein, described training data includes sample graph image set and each sample graph concentrated with described sample image As the corresponding mask information of difference, described mask information is used for indicating the position of the shadow region in corresponding sample image；With And carry out neural metwork training to obtain described full convolutional network using described training data.

In one embodiment, described computer program instructions make when being run by computer performed by described computer The step obtaining training data includes：Obtain initial image set, wherein, initial pictures and described sample that described initial pictures are concentrated Sample image in this image set corresponds；Each initial pictures concentrated for described initial pictures, generate predetermined number Shadow region；Each initial pictures concentrated for described initial pictures, by the moon of this initial pictures and described predetermined number Shadow zone domain is superimposed, to obtain the sample image corresponding with this initial pictures；And described initial pictures are concentrated Each initial pictures, superposed positions in this initial pictures for the shadow region according to described predetermined number obtain initial with this The corresponding mask information of sample image corresponding to image.

In one embodiment, described computer program instructions make when being run by computer performed by described computer The step generating the shadow region of predetermined number includes：Generate multiple images block；Number is randomly choosed from the plurality of image block Mesh is equal to the image block of described predetermined number；And the pixel value by the pixel in each image block in selected image block It is randomly set to the value in the range of default shading value, to obtain the shadow region of described predetermined number.

In one embodiment, make when being run by computer performed by described computer in described computer program instructions The step carrying out neural metwork training to obtain described full convolutional network using described training data before, described computer journey Sequence instruction also makes described computer execute when being run by computer：The size contracting of the sample image that described sample image is concentrated Put as standard size.

In one embodiment, described computer program instructions make when being run by computer performed by described computer The size scaling of the sample image that described sample image is concentrated includes for standard-sized step：For described sample graph image set In each sample image, keep this sample image the ratio of width to height constant while, by the height of this sample image and width In the greater zoom to normal size.

In one embodiment, described computer program instructions make when being run by computer performed by described computer The step obtaining training data includes：Receive described sample graph image set and each sample image with described sample image concentration divides Not corresponding labeled data, wherein, described labeled data includes the wheel for indicating the shadow region in corresponding sample image Wide outline data；And according to the outline data generation corresponding with each sample image that described sample image is concentrated and be somebody's turn to do The corresponding mask information of sample image.

In one embodiment, make when being run by computer performed by described computer in described computer program instructions The step described altimetric image to be checked being processed using full convolutional network before, described computer program instructions are being calculated Machine also makes described computer execution when running：The size scaling of described altimetric image to be checked is standard size.

In one embodiment, described computer program instructions make when being run by computer performed by described computer The size scaling of described altimetric image to be checked is included for standard-sized step：In the ratio of width to height keeping described altimetric image to be checked not While change, the greater in the height and width of described altimetric image to be checked is zoomed to normal size.

In one embodiment, described computer program instructions make when being run by computer performed by described computer Described altimetric image to be checked is carried out process using full convolutional network and include：Will be described for standard-sized described altimetric image input to be checked Full convolutional network, to obtain the primary knot of the positional information with regard to the shadow region in standard-sized described altimetric image to be checked Really；And the scaling according to described altimetric image to be checked and described Primary Outcome obtain described testing result.

Each module in image shadow Detection system according to embodiments of the present invention can be by according to embodiments of the present invention The processor of the electronic equipment of enforcement image shadow Detection run the computer program instructions that store in memory to realize, Or the computer that can store in the computer-readable recording medium of computer program according to embodiments of the present invention Instruction is realized when being run by computer.

Image shadow detection method according to embodiments of the present invention and device, using the full convolutional network training, permissible Direct detection shade whether there is, and can accurately be partitioned into shade place region, therefore the method have high precision, Adaptable feature, can be greatly enhanced the precision of image recognition in associated picture identification mission, reliability.Additionally, Image shadow detection method according to embodiments of the present invention and device also have the characteristics that processing speed is fast, model small volume, because This can easily be deployed on the mobile device of smart mobile phone, panel computer etc..

Although here by reference to Description of Drawings example embodiment it should be understood that above-mentioned example embodiment is merely exemplary , and be not intended to limit the scope of the invention to this.Those of ordinary skill in the art can carry out various changes wherein And modification, it is made without departing from the scope of the present invention and spirit.All such changes and modifications are intended to be included in claims Within required the scope of the present invention.

It should be understood that disclosed equipment and method in several embodiments provided herein, can be passed through it Its mode is realized.For example, apparatus embodiments described above are only schematically, for example, the division of described unit, and only It is only a kind of division of logic function, actual can have other dividing mode when realizing, and for example multiple units or assembly can be tied Close or be desirably integrated into another equipment, or some features can be ignored, or do not execute.

In specification mentioned herein, illustrate a large amount of details.It is to be appreciated, however, that the enforcement of the present invention Example can be put into practice in the case of not having these details.In some instances, known method, structure are not been shown in detail And technology, so as not to obscure the understanding of this description.

Similarly it will be appreciated that in order to simplify the present invention and help understand one or more of each inventive aspect, In description to the exemplary embodiment of the present invention, each feature of the present invention be sometimes grouped together into single embodiment, figure, Or in descriptions thereof.However, this method of the present invention should be construed to reflect following intention：I.e. required for protection Application claims more features than the feature being expressly recited in each claim.More precisely, weighing as corresponding As sharp claim is reflected, its inventive point is can be with the spy of all features of embodiment single disclosed in certain Levy to solve corresponding technical problem.Therefore, it then follows it is concrete that claims of specific embodiment are thus expressly incorporated in this Embodiment, wherein each claim itself is as the separate embodiments of the present invention.

It will be understood to those skilled in the art that in addition to mutually exclusive between feature, any combinations pair can be adopted All features disclosed in this specification (including adjoint claim, summary and accompanying drawing) and so disclosed any method Or all processes of equipment or unit are combined.Unless expressly stated otherwise, (including adjoint right will for this specification Ask, make a summary and accompanying drawing) disclosed in each feature can be replaced by the alternative features providing identical, equivalent or similar purpose.

Although additionally, it will be appreciated by those of skill in the art that some embodiments described herein include other embodiments In included some features rather than further feature, but the combination of the feature of different embodiment means to be in the present invention's Within the scope of and form different embodiments.For example, in detail in the claims, embodiment required for protection one of arbitrarily Can in any combination mode using.

The all parts embodiment of the present invention can be realized with hardware, or to run on one or more processor Software module realize, or with combinations thereof realize.It will be understood by those of skill in the art that can use in practice Microprocessor or digital signal processor (DSP) are realizing in image shadow Detection device according to embodiments of the present invention The some or all functions of a little modules.The present invention be also implemented as a part for executing method as described herein or The whole program of device of person (for example, computer program and computer program).Such program realizing the present invention is permissible Storage on a computer-readable medium, or can have the form of one or more signal.Such signal can from because Download on special net website and obtain, or provide on carrier signal, or provided with any other form.

It should be noted that above-described embodiment the present invention will be described rather than limits the invention, and ability Field technique personnel can design alternative embodiment without departing from the scope of the appended claims.In the claims, Any reference symbol between bracket should not be configured to limitations on claims.Word "comprising" does not exclude the presence of not Element listed in the claims or step.Word "a" or "an" before element does not exclude the presence of multiple such Element.The present invention can come real by means of the hardware including some different elements and by means of properly programmed computer Existing.If in the unit claim listing equipment for drying, several in these devices can be by same hardware branch To embody.The use of word first, second, and third does not indicate that any order.These words can be explained and run after fame Claim.

The above, the only specific embodiment of the present invention or the explanation to specific embodiment, the protection of the present invention Scope is not limited thereto, any those familiar with the art the invention discloses technical scope in, can be easily Expect change or replacement, all should be included within the scope of the present invention.Protection scope of the present invention should be with claim Protection domain is defined.

Claims

1. a kind of image shadow detection method, including：

Obtain altimetric image to be checked；And

Using full convolutional network, described altimetric image to be checked is processed, to obtain with regard to the shadow region in described altimetric image to be checked The testing result of the positional information in domain.

2. image shadow detection method as claimed in claim 1, wherein, described testing result includes shade probability graph, described The pixel value of each pixel in shade probability graph represents in described altimetric image to be checked and this pixel coordinate identical pixel There is the probability of shade in place.

3. image shadow detection method as claimed in claim 1 or 2, wherein, described image shadow detection method also includes：

Obtain training data, wherein, described training data includes sample graph image set and each sample concentrated with described sample image This image corresponding mask information respectively, described mask information is used for indicating the position of the shadow region in corresponding sample image Put；And

Carry out neural metwork training using described training data to obtain described full convolutional network.

4. image shadow detection method as claimed in claim 3, wherein, described mask information is and corresponding sample image chi Very little identical bianry image, sitting with the pixel within the shadow region being located at corresponding sample image in described bianry image Mark identical pixel have in the first pixel value, and, described bianry image be located at corresponding sample image shadow region The pixel coordinate identical pixel in overseas portion has the second pixel value.

5. image shadow detection method as claimed in claim 3, wherein, described acquisition training data includes：

Obtain initial image set, wherein, initial pictures and the sample graph of described sample image concentration that described initial pictures are concentrated As corresponding；

Each initial pictures that described initial pictures are concentrated,

Generate the shadow region of predetermined number；

The shadow region of this initial pictures and described predetermined number is superimposed, corresponding with this initial pictures to obtain Sample image；And

Superposed positions in this initial pictures for the shadow region according to described predetermined number obtain with corresponding to this initial pictures The corresponding mask information of sample image.

6. image shadow detection method as claimed in claim 5, wherein, the described shadow region generating predetermined number includes：

Generate multiple images block；

The image block that number is equal to described predetermined number is randomly choosed from the plurality of image block；And

The pixel value of the pixel in each image block in selected image block is randomly set in the range of default shading value Value, to obtain the shadow region of described predetermined number.

7. image shadow detection method as claimed in claim 3, wherein, carries out nerve net described using described training data Before network training is to obtain described full convolutional network, described image shadow detection method also includes：

The size scaling of the sample image that described sample image is concentrated is standard size.

8. image shadow detection method as claimed in claim 7, wherein, the described sample image that described sample image is concentrated Size scaling include for standard size：

Each sample image that described sample image is concentrated, while the ratio of width to height keeping this sample image is constant, will The greater in the height and width of this sample image zooms to normal size.

9. image shadow detection method as claimed in claim 3, wherein, described acquisition training data includes：

Receive described sample graph image set and each sample image corresponding labeled data respectively concentrated with described sample image, its In, described labeled data includes the outline data of the profile for indicating the shadow region in corresponding sample image；And

Generated relative with this sample image according to the corresponding outline data of each sample image concentrated with described sample image The mask information answered.

10. image shadow detection method as claimed in claim 1 or 2, wherein, is treated to described using full convolutional network described Before detection image is processed, described image shadow detection method also includes：

The size scaling of described altimetric image to be checked is standard size.

11. image shadow detection methods as claimed in claim 10, wherein, the described size scaling by described altimetric image to be checked Include for standard size：

Keep described altimetric image to be checked the ratio of width to height constant while, by the height of described altimetric image to be checked and width relatively Big person zooms to normal size.

12. image shadow detection methods as claimed in claim 10, wherein, described using full convolutional network to described to be detected Image carries out processing inclusion：

Standard-sized described altimetric image to be checked is inputted described full convolutional network, to obtain with regard to standard-sized described to be checked The Primary Outcome of the positional information of the shadow region in altimetric image；And

Scaling according to described altimetric image to be checked and described Primary Outcome obtain described testing result.

13. image shadow detection methods as claimed in claim 1, wherein, described full convolutional network is configured to be had with off line Network structure：Input layer, followed by two convolutional layers, connects a maximum pond layer, followed by two convolutional layers, connects one Individual maximum pond layer, followed by three convolutional layers, connects a maximum pond layer, followed by three convolutional layers, connects one Individual maximum pond layer, followed by three convolutional layers.

A kind of 14. image shadow Detection devices, including：

Image collection module, for obtaining altimetric image to be checked；And

Processing module, for being processed to described altimetric image to be checked using full convolutional network, to obtain with regard to described to be detected The testing result of the positional information of the shadow region in image.

15. image shadow Detection devices as claimed in claim 14, wherein, described testing result includes shade probability graph, institute The pixel value stating each pixel in shade probability graph represents in described altimetric image to be checked and this pixel coordinate identical picture There is the probability of shade in plain place.

The 16. image shadow Detection devices as described in claims 14 or 15, wherein, described image shadow Detection device also wraps Include：

Data acquisition module, for obtaining training data, wherein, described training data include sample graph image set and with described sample Each sample image in image set corresponding mask information respectively, described mask information is used for indicating in corresponding sample image Shadow region position；And

Training module, for carrying out neural metwork training to obtain described full convolutional network using described training data.

17. image shadow Detection devices as claimed in claim 16, wherein, described mask information is and corresponding sample image Equivalently-sized bianry image, in described bianry image be located at corresponding sample image shadow region within pixel Coordinate identical pixel have in the first pixel value, and, described bianry image be located at corresponding sample image shade The pixel coordinate identical pixel of region exterior has the second pixel value.

18. image shadow Detection devices as claimed in claim 16, wherein, described data acquisition module includes：

Initial pictures acquisition submodule, for obtaining initial image set, wherein, initial pictures and institute that described initial pictures are concentrated The sample image stating sample image concentration corresponds；

Shadow region generates submodule, for each initial pictures concentrated for described initial pictures, generates predetermined number Shadow region；

Superposition submodule, for each initial pictures concentrated for described initial pictures, this initial pictures are predetermined with described The shadow region of number is superimposed, to obtain the sample image corresponding with this initial pictures；And

Mask information obtains submodule, for each initial pictures concentrated for described initial pictures, according to described predetermined number Superposed positions in this initial pictures for the purpose shadow region obtain corresponding with the sample image corresponding to this initial pictures Mask information.

19. image shadow Detection devices as claimed in claim 18, wherein, described shadow region generates submodule and includes：

Image block signal generating unit, for generating multiple images block；

Image block select unit, for randomly choosing the image that number is equal to described predetermined number from the plurality of image block Block；And

Pixel value setup unit, for being randomly provided the pixel value of the pixel in each image block in selected image block For presetting the value in the range of shading value, to obtain the shadow region of described predetermined number.

20. image shadow Detection devices as claimed in claim 16, wherein, described image shadow Detection device also includes：

First Zoom module, described to obtain for carrying out neural metwork training in described training module using described training data Before full convolutional network, the size scaling of the sample image that described sample image is concentrated is standard size.

21. image shadow Detection devices as claimed in claim 20, wherein, described first Zoom module includes：

First scaling submodule, for each sample image concentrated for described sample image, is keeping this sample image While the ratio of width to height is constant, the greater in the height and width of this sample image is zoomed to normal size.

22. image shadow Detection devices as claimed in claim 16, wherein, described data acquisition module includes：

Receiving submodule, for receive described sample graph image set and with described sample image concentrate each sample image right respectively The labeled data answered, wherein, described labeled data includes the profile for indicating the shadow region in corresponding sample image Outline data；And

Mask information generates submodule, for according to the number of contours corresponding with each sample image that described sample image is concentrated According to the generation mask information corresponding with this sample image.

The 23. image shadow Detection devices as described in claims 14 or 15, wherein, described image shadow Detection device also wraps Include：

Second Zoom module, for carrying out processing it to described altimetric image to be checked using full convolutional network in described processing module Before, the size scaling of described altimetric image to be checked is standard size.

24. image shadow Detection devices as claimed in claim 23, wherein, described second Zoom module includes：

Second scaling submodule, for while the ratio of width to height keeping described altimetric image to be checked is constant, by described mapping to be checked The greater in the height and width of picture zooms to normal size.

25. image shadow Detection devices as claimed in claim 23, wherein, described processing module includes：

Input submodule, for standard-sized described altimetric image to be checked is inputted described full convolutional network, to obtain with regard to mark The Primary Outcome of the positional information of shadow region in the altimetric image described to be checked of object staff cun；And

Testing result obtains submodule, for described in the scaling according to described altimetric image to be checked and the acquisition of described Primary Outcome Testing result.

26. image shadow Detection devices as claimed in claim 14, wherein, described full convolutional network is configured to be had with off line Network structure：Input layer, followed by two convolutional layers, connects a maximum pond layer, followed by two convolutional layers, connects one Individual maximum pond layer, followed by three convolutional layers, connects a maximum pond layer, followed by three convolutional layers, connects one Individual maximum pond layer, followed by three convolutional layers.