CN108564587A - A kind of a wide range of remote sensing image semantic segmentation method based on full convolutional neural networks - Google Patents
A kind of a wide range of remote sensing image semantic segmentation method based on full convolutional neural networks Download PDFInfo
- Publication number
- CN108564587A CN108564587A CN201810186195.1A CN201810186195A CN108564587A CN 108564587 A CN108564587 A CN 108564587A CN 201810186195 A CN201810186195 A CN 201810186195A CN 108564587 A CN108564587 A CN 108564587A
- Authority
- CN
- China
- Prior art keywords
- image
- remote sensing
- sensing image
- semantic segmentation
- wide range
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/90—Determination of colour characteristics
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10032—Satellite or aerial image; Remote sensing
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Analysis (AREA)
Abstract
The a wide range of remote sensing image semantic segmentation method based on full convolutional neural networks that the invention discloses a kind of, including data mark, model training, prediction result three phases；The features such as this method is for remote sensing image high-precision, a wide range of, multispectral information, the pretreatment operation of wave band synthesis, visual fusion, image cutting has been carried out to remote sensing image, the abundant degree of the skill exptended sample of data enhancing has been used training set, if the data set of tape label can be got, first model can be trained with it, the initialization for carrying out target data model training, to reduce the workload manually marked.In order to improve the accuracy rate of result, the grid cutting that the method for the present invention has carried out having overlapping to image after being predicted respectively, splices prediction result image in order, and carry out medium filtering and reduce rough part in noise and image, finally achieve higher accuracy rate.
Description
Technical field
The invention belongs to remote sensing image identification and depth learning technology fields, and in particular to one kind being based on full convolutional Neural net
The a wide range of remote sensing image semantic segmentation method of network.
Background technology
Remote sensing is to carry out the imaging of specific electromagnetic wave spectral coverage to the earth by the sensor on satellite, be with aeroplane photography
The a special kind of skill to grow up based on technology；It, can in a short time, to regional on a large scale on the earth by remote sensing
Multi-level, various visual angles observations are carried out, are the important means for obtaining environmental information and earth resource.
Remote sensing technology is to weigh the one of the important signs that of national a scientific and technological level and comprehensive strength, and China is always very
Pay attention to the development of remote sensing technology so that remote sensing technology is rapidly developed；Currently, remote sensing technology has been widely used Yu Haiyang, gas
As various fields such as, agricultural, military affairs, forestry, water resource, GEOLOGICAL ENVIRONMENT SURVEY, environmental protection, land use, urban plannings.
In face of so open application prospect, the information extraction technology of remote sensing image is particularly important；It can be said that remote sensing
Final goal be just to be able to extract useful information from image, get knowledge.Most basic remote sensing image letter
Breath extraction has the classification and identification of object, under huge data volume, is manually classified and identifies obviously no longer feasible, needed
More intelligent means；On this basis, remote sensing image can just be made to obtain more wider array of applications.In this respect, with meter
The development of calculation machine vision and depth learning technology, more and more methods and model emerge in large numbers, and are carried for the information extraction of remote sensing image
Strong means are supplied.
Currently, high score No.1 (GF-1) to 9 satellites of high score nine (GF-9) have succeeded in sending up, cover panchromatic, more
Numerous imaging types such as spectrum, EO-1 hyperion, radar.Wherein, it is more representational have be equipped with panchromatic and multispectral camera
No. two satellites of high score No.1 satellite and high score are equipped with C frequency ranges multi polarized SAR (Synthetic Aperture
Radar, SAR) No. three satellites of high score, be equipped with No. five satellites of high score of EO-1 hyperion camera.
Multispectral image generally has 4 wave spectrum frequency ranges, typically blue, green, red and four wave bands of near-infrared；SAR
Only there are one channel, image is generally gray level image for imaging, is chiefly used in terms of detecting landform；Hyperspectral imaging is on electromagnetic spectrum
To target area while being imaged with multiple spectral bands continuously segmented, include ultraviolet, visible-range, near-infrared and in it is red
Outer equal regions.
Highest resolution ratio is No. two satellites of high score, equipped with 1 meter of panchromatic and 4 meters of multispectral camera of high-resolution, is
The civilian Optical remote satellite and first, China satellite for reaching sub-meter grade spatial resolution of China's independent development are high
Positioning accuracy and quick attitude maneuver ability, greatly improve satellite INTEGRATED SIGHT efficiency, once reached advanced world standards,
Application for fields such as the fine-grained management in city, traffic programme, all kinds of resource investigations and monitorings provides supporting.
Image, semantic segmentation is the concept in computer vision, refers to, according to color, shape and Texture eigenvalue, to scheme
Region as being divided into multiple non-overlapping copies makes the part for being divided into the same region have similar feature, without same district
Feature between domain has more apparent difference, be it is a kind of divide the image into multiple isolated areas with respective characteristic, and
Extract the technology of interesting target.
In high score satellite remote-sensing image, the specific objective in image can be divided by image, semantic cutting techniques
It is marked, the specific information in remote sensing image, such as identification and division, the identification of road network, the vegetation point in house is extracted with this
From etc.；These are all the information extractions on remote sensing image basis, and other applications but more need to carry out on this basis,
And the accuracy rate of the information extraction on basis, directly influence subsequent use.
Invention content
In view of above-mentioned, the present invention provides a kind of a wide range of remote sensing image semantic segmentation side based on full convolutional neural networks
Method can realize a wide range of remote sensing image semantic segmentation of efficiently and accurately.
A kind of a wide range of remote sensing image semantic segmentation method based on full convolutional neural networks, includes the following steps：
(1) high-resolution remote sensing image is obtained, visual fusion and wave band synthetic operation are carried out to it；
(2) different classes of object is marked in the remote sensing image after visual fusion and wave band synthesis, obtains label figure
Picture；
(3) full convolutional neural networks model is designed, and it is trained and is joined using remote sensing image and tag image
Number fine tuning；
(4) grid for needing the remote sensing image of semantic segmentation have overlapping is sliced, and then each sectioning image is defeated
Enter into full convolutional neural networks model output to obtain corresponding single channel gray level image and divide prediction result；
(5) finally obtained all segmentation prediction result images are orderly spliced and is post-processed.
Further, the step (1) the specific implementation process is as follows：
1.1 take gray level image that corresponding wavelength size in remote sensing image is first three wave band respectively as the B in RGB image
Channel image, G channel images and R channel images, and then synthesize a RGB color image；
Above-mentioned RGB color image is transformed into IHS coloured images by 1.2, and be amplified to it is panchromatic in remote sensing image
Channel gray level image is onesize；
1.3 by the I channel images in IHS coloured images after the gray level image of the panchromatic channel and amplification into column hisgram
Matching, and the result after matching is substituted into the I channel images in IHS coloured images；
IHS coloured images after the completion of replacement are converted back RGB color image by 1.4, that is, complete visual fusion and wave band
Synthetic operation.
Further, the step (2) the specific implementation process is as follows：
The RGB color image of remote sensing image after visual fusion and wave band synthesis is cut into multiple 1500 × 1500 greatly by 2.1
Small segment；
2.2 for any segment, and the form for retouching side with polygon using image labeling tool will be different classes of in segment
Object mark out, and preserve the vertex position information and object generic information of polygon；
Segment is converted into single pass gray level image by 2.3 according to classification information, which is tag image,
In the gray value of each pixel be the object type indicated belonging to pixel, 0 represents background, and 1,2,3 ..., N correspondences represent difference
Object type, N is the total categorical measure of all objects in remote sensing image.
Further, the structure of the full convolutional neural networks model from be input to output successively by 4 encrypting modules and 4
A deciphering module connection composition, the encrypting module is from output is input to successively by 5 convolution combination layers and 1 pond layer connection
Composition, from being input to, output up-samples layer to the deciphering module by 1 successively and 5 convolution combination layers connections form, the volume
Product combination layer is connected with 1 active coating and is formed by 1 convolutional layer, 1 batch normalization layer successively from being input to output；The convolution
Convolution kernel size in layer is 3 × 3, step-length 1, back gauge 1, and the active coating samples ReLU functions, times of the pond layer
Number is 2, and the multiple of the up-sampling layer is 2；Convolution kernel number in 4 encrypting modules from be input to output be followed successively by 64,
128,256,512, the convolution kernel number in 4 deciphering modules is followed successively by 512,256,128,64 from output is input to.
Further, the full convolutional neural networks model in the step (3) receives the coloured image that size is n × n and makees
For input, output is then the single channel gray level image of m × m sizes as segmentation prediction result；In training, from remote sensing image
The colored segment that size is n × n is arbitrarily intercepted in RGB color image, make corresponding position m in the colour segment and tag image ×
The gray scale segment of m sizes is input to as one group of sample in full convolutional neural networks model, and multigroup sample inputs one by one to be learned repeatedly
Model training task can be completed in habit, and n and m are natural number of the setting more than 1.
Further, the grid for needing the remote sensing image of semantic segmentation have overlapping is sliced in the step (4), i.e.,
Ensure that the overlapping back gauge of two neighboring sectioning image is k picture when carrying out grid slice to the RGB color image of the remote sensing image
The size of element, sectioning image is n × n, and k=n-m+2p, p are natural number of the setting more than 1.
Further, the step (5) the specific implementation process is as follows：
5.1 be m × m for the corresponding segmentation prediction result image of any sectioning image, size, crops bottom left thereon
The marginal portion of right each p/2 pixel wide；
5.2 splice the segmentation prediction result image after all cuttings in order, obtain a complete remote sensing image
Semantic segmentation result images；
5.3 finally carry out the medium filtering that core size is 5 × 5 to the semantic segmentation result images.
Based on the above-mentioned technical proposal, the present invention has following advantageous effects：
(1) the big, feature with high accuracy of the invention for remote sensing image range, a wide range of remote sensing image can be handled by devising
Using deep learning disaggregated model image, semantic divide flow.
(2) present invention utilizes the network structure of deep layer, and the more general object features of extraction can be in different remote sensing images
It is all well used on data set.
(3) network model proposed by the present invention can be trained as form class on public data collection, then compared with
Model fine tuning is carried out on small-scale artificial labeled data, reduces labor workload.
Description of the drawings
Fig. 1 is the techniqueflow schematic diagram of the method for the present invention.
Fig. 2 is the network architecture schematic diagram in the method for the present invention.
Specific implementation mode
In order to more specifically describe the present invention, below in conjunction with the accompanying drawings and specific implementation mode is to technical scheme of the present invention
It is described in detail.
As shown in Figure 1, the present invention is based on a wide range of remote sensing image semantic segmentation methods of full convolutional neural networks, including such as
Lower step：
(1) Remote Sensing Image Fusion is synthesized with wave band.
In the training of machine learning, the quality and quantity of sample are all very important；In practical application scene, sample
This quantity can usually be met by some skills, but the quality of sample then places one's entire reliance upon artificial mark.Due to being artificial
Mark, so including often a certain number of data not marked correctly in training, such data will necessarily
Some negative effects are generated to final result；Negatively affected caused by order to reduce this data mark as far as possible, need into
Row Remote Sensing Image Fusion is synthesized with wave band, is made it easier to accurately be labeled, is as follows：
Panchromatic, the multi-spectral remote sensing image that 1.1 satellites of high score two provide.The text of full color spectrum image and multispectral image
Part is stored separately, and is different from the storage mode of the RGB channel and 256 color ranges of normal image；Only there are one logical for full color spectrum image
Road, color range 1000；It is equally 1000 that multispectral image, which then has 4 channels, color range,.Therefore, image can not directly display, and need
By four channel separations of multispectral image, blue, green, red three channels therein are chosen, by all pixels in these three channels
Gray value carry out linear stretch, be scaled to 256 color ranges, later merge three channel combined color images.
Coloured image is transformed into IHS color systems by 1.2 from RGB color system, and wherein I (Intensity) represents lightness,
H (Hue) represents color code, and S (Saturation) represents saturation degree, and being transformed into HIS color systems from RGB color system generally uses
Following triangular transformation formula：
The height and width of 1.3 full-colour images are about 4 times of coloured image, and coloured image is amplified to and panchromatic gray scale
Image is onesize.
The gray value of full-colour image is mainly interpreted as image by 1.4 Remote Sensing Image Fusions based on IHS transformation
Brightness, therefore the channels I of the gray level image of panchromatic channel and coloured image are subjected to Histogram Matching, replace the channels I of script.
Image is transformed into RGB color system by 1.5 from IHS color systems, obtains the coloured image of increase resolution.
(2) mark part remote sensing image is as training data.
A scape chromatic image size after fusion is generally more than 29000 × 27000, and file size is more than 1GB, this size
It is unfavorable for the mark of image, needs to carry out cutting to image, the image of synthesis can be cut by 1500 × 1500 sizes more
Open image.
The data mark of semantic segmentation only need to give each image unlike the data of simple image classification problem mark
Mark classification, but must mark what classification each pixel in image belongs to, generally require by some tools and
Program is completed.In view of facilitating modification and storage, side can be retouched by polygon to choose a part for image, and more thus
Pixel in the shape of side specifies item name, and each vertex of one group of polygon is recorded, the title of generic is enclosed, and leads to
It crosses specific format to be stored in other file, generally be stored using xml or json formats, and closed surely in each document
The path of the image file of connection.
The data note of mark needs to input disaggregated model together with original image in the form of images, generally uses a gray scale
Image, the gray value of each pixel indicate the classification belonging to the pixel；The text text of polygon vertex information will be had recorded
Part generates label images, needs to create a gray level image, different classes of pixel is set as different gray values, i.e., each
The gray value of pixel indicates the type belonging to the pixel, and 0 is background, and continuous increase indicates different objects respectively since 1
Type.
(3) network model design and training.
The present invention devises the mould based on full convolutional neural networks with reference to the network model of the vgg-16 versions of SegNet
Type structure is divided into 4 Encoder stages and corresponding 4 Decoder stages and last Softmax layers, as shown in Fig. 2,
Each Encoder stage includes 5 groups of convolution combination layers and one layer of pond layer, each Decoder stage includes to be adopted on one layer
Sample layer and 5 groups of convolution combination layers, each group of convolution combination layer are respectively convolutional layer in order, batch normalization layer, use ReLU letters
The concrete structure of several active coatings, model is as follows：
The Encoder stages 1：It is 3 × 3 convolution kernel that convolutional layer in each group of convolution combination layer, which all uses 64 sizes,
Padding is not set, the convolution operation that step-length is 1 is carried out；After carrying out 5 groups of convolution operations, the height and width of image can reduce 10
A pixel then carries out the operation of maximum value pondization, the height of image and width is all reduced 2 times.
The Encoder stages 2：It is 3 × 3 convolution kernel that convolutional layer in each group of convolution combination layer, which all uses 128 sizes,
Padding is 1, carries out the convolution operation that step-length is 1, keeps image size constant in convolution process, then carry out maximum value pond
Change operation, the height of image and width are all reduced 2 times.
The Encoder stages 3：It is 3 × 3 convolution kernel that convolutional layer in each group of convolution combination layer, which all uses 256 sizes,
Padding is 1, carries out the convolution operation that step-length is 1, keeps image size constant in convolution process, then carry out maximum value pond
Change operation, the height of image and width are all reduced 2 times.
The Encoder stages 4：It is 3 × 3 convolution kernel that convolutional layer in each group of convolution combination layer, which all uses 512 sizes,
Padding is 1, carries out the convolution operation that step-length is 1, keeps image size constant in convolution process, then carry out maximum value pond
Change operation, the height of image and width are all reduced 2 times.
The Decoder stages 4：First up-sampled the pond layer that 2 times of image magnification is restored to the corresponding Encoder stages
Size before, the convolution operation in each later group of convolution combination layer are identical as the Encoder stages 4.
The Decoder stages 3：It is first up-sampled 2 times of image magnification, the volume operation in each later group of convolution combination layer
It is identical as the Encoder stages 3.
The Decoder stages 2：It is first up-sampled 2 times of image magnification, the convolution behaviour in each later group of convolution combination layer
Make identical as the Encoder stages 2.
The Decoder stages 1：It is first up-sampled 2 times of image magnification, image is restored to the progress of Encoder stages 1 at this time
Size before pondization operation, 10 pixels all fewer than the height of original image and width；In each later group of convolution combination layer
Convolutional layer all to use 64 sizes be 3 × 3 convolution kernel, do not set padding, carry out the convolution operation that step-length is 1, complete 5
After group convolution operation, the height and width of image reduce by 10 pixels again.
It it is finally Softmax layers, the sum of the value by each pixel in different dimensions is normalized, by the value table of each dimension
It is shown as the probability that the pixel belongs to this classification.
The image size that network receives is 266 × 266, and the prediction result image size of output is 246 × 246, in training
When can to training set carry out data enhancing, EDS extended data set, mainly include several measures：
● rotation/affine transformation：By the certain angle of image rotation, or carry out affine transformation.
● it is turning-over changed：Image is subjected to horizontal or vertical overturning.
● scale transformation：The size of image is zoomed in or out by a certain percentage.
● noise disturbance：Noise is added in the picture.
● translation transformation：Image is translated into a distance by some direction.
● random interception：The part of random interception image.
If fairly large tagged data collection can be obtained, can first be used for model is trained, then with train
Good model is initialized, and is trained on target data set, is finely adjusted to model again, can be reduced so artificial
The work of mark.
(4) overlapping grid slice has been carried out to complete remote sensing image.
Network model is had as defined in size to the image data of input, so to carry out language to large-scale remote sensing image
Justice segmentation needs image carrying out grid slice；Simple grid slice can cause the marginal portion of each sectioning image to be predicted
It is inaccurate, it is therefore desirable to prediction result image be cut, cast out and predict inaccurate marginal portion.After making cutting
Result images between can completely splice, need when carrying out grid slice to complete remote sensing image, to allow adjacent cut
Piece overlaps.
After model prediction, the image of output width fewer than input picture is the edge of 10 pixels, it is also necessary to
Crop the edge of 30 pixels.Then the image inputted is 266 × 266, the 186 of part centered on corresponding result images ×
The image of 186 sizes.Therefore adjacent slice needs the lap that width is 80 pixels.
(5) semantic segmentation prediction is carried out to each sectioning image.
By each size be 266 × 266 sectioning image input network model, can obtain size be 246 × 246 it is pre-
Result images are surveyed, which is a single channel gray-scale map, and the gray value of each pixel represents corresponding classification.
(6) result images after prediction are orderly spliced.
Prediction result image after conversion is subjected to orderly splicing by row by row, obtains complete remote sensing image semanteme point
Cut result.
(7) result images are post-processed.
The result images of prediction have had a higher accuracy rate, but it sometimes appear that the case where edge out-of-flatness, so can
To carry out image smoothing operation in prediction result, result is optimized.It is big that the present invention has carried out core to prediction result image
The small median filtering operation for being 5 × 5, medium filtering are the median works that grey scale pixel value in a certain range is chosen to each pixel
For the value of the pixel, what is represented due to the gray value of prediction result image is type, so new gray value cannot be generated, is used
Medium filtering would not encounter this problem；It, can be with by the way that a kind of corresponding color is arranged to each classification in advance after having handled
By prediction result gray level image, corresponding coloured image is converted to.If image is only divided into two classes, can use respectively black
Color and white indicate.
The above-mentioned description to embodiment can be understood and applied the invention for ease of those skilled in the art.
Person skilled in the art obviously easily can make various modifications to above-described embodiment, and described herein general
Principle is applied in other embodiment without having to go through creative labor.Therefore, the present invention is not limited to the above embodiments, ability
Field technique personnel announcement according to the present invention, the improvement made for the present invention and modification all should be in protection scope of the present invention
Within.
Claims (7)
1. a kind of a wide range of remote sensing image semantic segmentation method based on full convolutional neural networks, includes the following steps：
(1) high-resolution remote sensing image is obtained, visual fusion and wave band synthetic operation are carried out to it；
(2) different classes of object is marked in the remote sensing image after visual fusion and wave band synthesis, obtains tag image；
(3) full convolutional neural networks model is designed, and it is trained using remote sensing image and tag image and parameter is micro-
It adjusts；
(4) grid for needing the remote sensing image of semantic segmentation have overlapping is sliced, and then each sectioning image is input to
Output obtains corresponding single channel gray level image and divides prediction result in full convolutional neural networks model；
(5) finally obtained all segmentation prediction result images are orderly spliced and is post-processed.
2. a wide range of remote sensing image semantic segmentation method according to claim 1, it is characterised in that：The step (1)
The specific implementation process is as follows：
1.1 take gray level image that corresponding wavelength size in remote sensing image is first three wave band respectively as the channel B in RGB image
Image, G channel images and R channel images, and then synthesize a RGB color image；
Above-mentioned RGB color image is transformed into IHS coloured images by 1.2, and is amplified to and the panchromatic channel in remote sensing image
Gray level image is onesize；
1.3 by the I channel images in IHS coloured images after the gray level image of the panchromatic channel and amplification into column hisgram
Match, and the result after matching is substituted into the I channel images in IHS coloured images；
IHS coloured images after the completion of replacement are converted back RGB color image by 1.4, that is, complete visual fusion and wave band synthesis
Operation.
3. a wide range of remote sensing image semantic segmentation method according to claim 1, it is characterised in that：The step (2)
The specific implementation process is as follows：
The RGB color image of remote sensing image after visual fusion and wave band synthesis is cut into multiple 1500 × 1500 sizes by 2.1
Segment；
2.2 for any segment, and the form on side is retouched in segment by different classes of object with polygon using image labeling tool
Body, which marks out, to be come, and preserves the vertex position information and object generic information of polygon；
Segment is converted into single pass gray level image by 2.3 according to classification information, which is tag image, wherein often
The gray value of a pixel is the object type indicated belonging to pixel, and 0 represents background, and 1,2,3 ..., N correspondences represent different objects
Body type, N are the categorical measure that all objects are total in remote sensing image.
4. a wide range of remote sensing image semantic segmentation method according to claim 1, it is characterised in that：The full convolutional Neural
The structure of network model is connected with 4 deciphering modules by 4 encrypting modules form successively from output is input to, the encrypting module
Be made of successively 5 convolution combination layers and 1 pond layer connection from output is input to, the deciphering module from be input to export according to
Secondary to be made of 1 up-sampling layer and 5 convolution combination layer connections, the convolution combination layer is from output is input to successively by 1 volume
Lamination, 1 batch of normalization layer connect composition with 1 active coating；Convolution kernel size in the convolutional layer is 3 × 3, step-length 1,
Back gauge is 1, and the active coating samples ReLU functions, and the multiple of the pond layer is 2, and the multiple of the up-sampling layer is 2；4
Convolution kernel number in encrypting module is followed successively by 64,128,256,512 from output is input to, the convolution kernel in 4 deciphering modules
Number is followed successively by 512,256,128,64 from output is input to.
5. a wide range of remote sensing image semantic segmentation method according to claim 1, it is characterised in that：In the step (3)
Full convolutional neural networks model receive size be n × n coloured image as input, output then be m × m sizes single channel
Gray level image is as segmentation prediction result；In training, it is n × n that size is arbitrarily intercepted from the RGB color image of remote sensing image
Colored segment, so that the gray scale segment of corresponding position m × m sizes in the colour segment and tag image is inputted as one group of sample
Into full convolutional neural networks model, multigroup sample inputs repetition learning and model training task can be completed one by one, and n and m are to set
Surely it is more than 1 natural number.
6. a wide range of remote sensing image semantic segmentation method according to claim 5, it is characterised in that：In the step (4)
The grid for needing the remote sensing image of semantic segmentation have overlapping is sliced, i.e., the RGB color image of the remote sensing image is carried out
Grid ensures that the overlapping back gauge of two neighboring sectioning image is k pixel when being sliced, the size of sectioning image is n × n, k=n-m
+ 2p, p are natural number of the setting more than 1.
7. a wide range of remote sensing image semantic segmentation method according to claim 6, it is characterised in that：The step (5)
The specific implementation process is as follows：
5.1 be m × m for the corresponding segmentation prediction result image of any sectioning image, size, and it is each up and down to crop it
The marginal portion of p/2 pixel wide；
5.2 splice the segmentation prediction result image after all cuttings in order, obtain the language of a complete remote sensing image
Adopted segmentation result image；
5.3 finally carry out the medium filtering that core size is 5 × 5 to the semantic segmentation result images.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810186195.1A CN108564587A (en) | 2018-03-07 | 2018-03-07 | A kind of a wide range of remote sensing image semantic segmentation method based on full convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810186195.1A CN108564587A (en) | 2018-03-07 | 2018-03-07 | A kind of a wide range of remote sensing image semantic segmentation method based on full convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108564587A true CN108564587A (en) | 2018-09-21 |
Family
ID=63531493
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810186195.1A Pending CN108564587A (en) | 2018-03-07 | 2018-03-07 | A kind of a wide range of remote sensing image semantic segmentation method based on full convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108564587A (en) |
Cited By (42)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108761843A (en) * | 2018-05-29 | 2018-11-06 | 杭州视氪科技有限公司 | A kind of blind person's auxiliary eyeglasses detected for the water surface and puddle |
| CN109215038A (en) * | 2018-09-29 | 2019-01-15 | 中国资源卫星应用中心 | A kind of intelligent information retrieval method and system based on remote sensing image |
| CN109241658A (en) * | 2018-09-27 | 2019-01-18 | 中国电子科技集团公司第五十四研究所 | Satellite dish Analysis method based on remote sensing image |
| CN109377479A (en) * | 2018-09-27 | 2019-02-22 | 中国电子科技集团公司第五十四研究所 | Satellite dish object detection method based on remote sensing image |
| CN109490861A (en) * | 2018-10-29 | 2019-03-19 | 北京科技大学 | A kind of blast-furnace line extracting method |
| CN109544468A (en) * | 2018-10-26 | 2019-03-29 | 浙江师范大学 | A kind of image data amplification method |
| CN109583302A (en) * | 2018-10-29 | 2019-04-05 | 杭州电子科技大学 | A kind of mobile robot area of feasible solutions training dataset extending method |
| CN109614973A (en) * | 2018-11-22 | 2019-04-12 | 华南农业大学 | Rice seedling and Weeds at seedling image, semantic dividing method, system, equipment and medium |
| CN109711413A (en) * | 2018-12-30 | 2019-05-03 | 陕西师范大学 | Image, semantic dividing method based on deep learning |
| CN109886238A (en) * | 2019-03-01 | 2019-06-14 | 湖北无垠智探科技发展有限公司 | Unmanned plane Image Change Detection algorithm based on semantic segmentation |
| CN110033045A (en) * | 2019-04-17 | 2019-07-19 | 内蒙古工业大学 | A kind of method and apparatus of trained identification image atomization |
| CN110197505A (en) * | 2019-05-30 | 2019-09-03 | 西安电子科技大学 | Remote sensing images binocular solid matching process based on depth network and semantic information |
| CN110207671A (en) * | 2018-12-29 | 2019-09-06 | 中国科学院软件研究所 | A kind of space-based intelligence imaging system |
| CN110210513A (en) * | 2019-04-23 | 2019-09-06 | 深圳信息职业技术学院 | Data classification method, device and terminal device |
| CN110245747A (en) * | 2019-06-21 | 2019-09-17 | 华中师范大学 | Image processing method and device based on full convolutional neural networks |
| CN110335245A (en) * | 2019-05-21 | 2019-10-15 | 青岛科技大学 | Cage netting damage monitoring method and system based on monocular space and time continuous image |
| CN110532224A (en) * | 2019-08-13 | 2019-12-03 | 武汉中海庭数据技术有限公司 | A kind of file management system and method for deep learning mark sample |
| CN110600105A (en) * | 2019-08-27 | 2019-12-20 | 武汉科技大学 | CT image data processing method, device and storage medium |
| CN110765853A (en) * | 2019-09-10 | 2020-02-07 | 中国科学院上海技术物理研究所 | Image processing method of multi-spectrum phase machine |
| CN111046835A (en) * | 2019-12-24 | 2020-04-21 | 杭州求是创新健康科技有限公司 | Eyeground illumination multiple disease detection system based on regional feature set neural network |
| CN111242228A (en) * | 2020-01-16 | 2020-06-05 | 武汉轻工大学 | Hyperspectral image classification method, device, equipment and storage medium |
| CN111311571A (en) * | 2020-02-13 | 2020-06-19 | 上海小萌科技有限公司 | Target information acquisition method, system, device and readable storage medium |
| CN111476170A (en) * | 2020-04-09 | 2020-07-31 | 首都师范大学 | Remote sensing image semantic segmentation method combining deep learning and random forest |
| CN111505738A (en) * | 2020-03-17 | 2020-08-07 | 上海眼控科技股份有限公司 | Method and equipment for predicting meteorological factors in numerical weather forecast |
| CN111582004A (en) * | 2019-02-15 | 2020-08-25 | 阿里巴巴集团控股有限公司 | Target area segmentation method and device in ground image |
| CN111597861A (en) * | 2019-02-21 | 2020-08-28 | 中科星图股份有限公司 | System and method for automatically interpreting ground object of remote sensing image |
| CN111814771A (en) * | 2020-09-04 | 2020-10-23 | 支付宝(杭州)信息技术有限公司 | Image processing method and device |
| CN111832569A (en) * | 2020-06-19 | 2020-10-27 | 西北大学 | High-spectrum classification and segmentation based mural paint layer shedding disease labeling method |
| CN111882563A (en) * | 2020-07-13 | 2020-11-03 | 浙江大学 | Semantic segmentation method based on directional convolutional network |
| CN112084923A (en) * | 2020-09-01 | 2020-12-15 | 西安电子科技大学 | Semantic segmentation method for remote sensing image, storage medium and computing device |
| CN112348820A (en) * | 2020-11-06 | 2021-02-09 | 天津市勘察设计院集团有限公司 | Remote sensing image semantic segmentation method based on depth discrimination enhancement network |
| CN112683921A (en) * | 2019-10-17 | 2021-04-20 | 神讯电脑(昆山)有限公司 | Image scanning method and image scanning system for metal surface |
| CN112784806A (en) * | 2021-02-04 | 2021-05-11 | 中国地质科学院矿产资源研究所 | Lithium-containing pegmatite vein extraction method based on full convolution neural network |
| CN113327250A (en) * | 2021-05-28 | 2021-08-31 | 深圳前海微众银行股份有限公司 | Multispectral image segmentation method and device, electronic device and storage medium |
| CN113468951A (en) * | 2021-05-20 | 2021-10-01 | 华东师范大学 | Method for detecting mangrove landform based on Landsat satellite remote sensing |
| CN113627292A (en) * | 2021-07-28 | 2021-11-09 | 广东海启星海洋科技有限公司 | Remote sensing image identification method and device based on converged network |
| WO2021238992A1 (en) * | 2020-05-26 | 2021-12-02 | 杭州海康威视数字技术股份有限公司 | Neural network training method and apparatus, electronic device, and readable storage medium |
| CN113820012A (en) * | 2020-06-19 | 2021-12-21 | 华为技术有限公司 | Multispectral image processing method and device |
| CN114898097A (en) * | 2022-06-01 | 2022-08-12 | 首都师范大学 | Image recognition method and system |
| CN115423829A (en) * | 2022-07-29 | 2022-12-02 | 江苏省水利科学研究院 | Method and system for rapidly extracting water body from single-band remote sensing image |
| CN115661680A (en) * | 2022-11-15 | 2023-01-31 | 北京轨道未来空间科技有限公司 | Satellite remote sensing image processing method |
| CN116883853A (en) * | 2023-01-12 | 2023-10-13 | 河南大学 | Crop space-time information remote sensing classification method based on transfer learning |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107153822A (en) * | 2017-05-19 | 2017-09-12 | 北京航空航天大学 | A kind of smart mask method of the semi-automatic image based on deep learning |
| CN107527352A (en) * | 2017-08-09 | 2017-12-29 | 中国电子科技集团公司第五十四研究所 | Remote sensing Ship Target contours segmentation and detection method based on deep learning FCN networks |
| CN107610141A (en) * | 2017-09-05 | 2018-01-19 | 华南理工大学 | A kind of remote sensing images semantic segmentation method based on deep learning |
-
2018
- 2018-03-07 CN CN201810186195.1A patent/CN108564587A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107153822A (en) * | 2017-05-19 | 2017-09-12 | 北京航空航天大学 | A kind of smart mask method of the semi-automatic image based on deep learning |
| CN107527352A (en) * | 2017-08-09 | 2017-12-29 | 中国电子科技集团公司第五十四研究所 | Remote sensing Ship Target contours segmentation and detection method based on deep learning FCN networks |
| CN107610141A (en) * | 2017-09-05 | 2018-01-19 | 华南理工大学 | A kind of remote sensing images semantic segmentation method based on deep learning |
Non-Patent Citations (4)
| Title |
|---|
| MICHAEL KAMPFFMEYER，ET AL.: "Semantic Segmentation of Small Objects and Modeling of Uncertainty in Urban Remote Sensing Images Using Deep Convolutional Neural Networks", 《2016 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION WORKSHOPS (CVPRW)》 * |
| VIJAY BADRINARAYANAN，ET AL.: "SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
| 文晓青: "多源图像融合算法在水利图像处理中的应用研究", 《中国优秀硕士学位论文全文数据库信息科技辑》 * |
| 高凯珺等: "基于深度学习的无人车夜视图像语义分割", 《应用光学》 * |
Cited By (58)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108761843A (en) * | 2018-05-29 | 2018-11-06 | 杭州视氪科技有限公司 | A kind of blind person's auxiliary eyeglasses detected for the water surface and puddle |
| CN109241658A (en) * | 2018-09-27 | 2019-01-18 | 中国电子科技集团公司第五十四研究所 | Satellite dish Analysis method based on remote sensing image |
| CN109377479A (en) * | 2018-09-27 | 2019-02-22 | 中国电子科技集团公司第五十四研究所 | Satellite dish object detection method based on remote sensing image |
| CN109215038A (en) * | 2018-09-29 | 2019-01-15 | 中国资源卫星应用中心 | A kind of intelligent information retrieval method and system based on remote sensing image |
| CN109544468A (en) * | 2018-10-26 | 2019-03-29 | 浙江师范大学 | A kind of image data amplification method |
| CN109544468B (en) * | 2018-10-26 | 2023-10-13 | 浙江师范大学 | Image data amplification method |
| CN109490861B (en) * | 2018-10-29 | 2020-06-02 | 北京科技大学 | Blast furnace burden line extraction method |
| CN109490861A (en) * | 2018-10-29 | 2019-03-19 | 北京科技大学 | A kind of blast-furnace line extracting method |
| CN109583302A (en) * | 2018-10-29 | 2019-04-05 | 杭州电子科技大学 | A kind of mobile robot area of feasible solutions training dataset extending method |
| CN109614973A (en) * | 2018-11-22 | 2019-04-12 | 华南农业大学 | Rice seedling and Weeds at seedling image, semantic dividing method, system, equipment and medium |
| CN110207671A (en) * | 2018-12-29 | 2019-09-06 | 中国科学院软件研究所 | A kind of space-based intelligence imaging system |
| CN109711413A (en) * | 2018-12-30 | 2019-05-03 | 陕西师范大学 | Image, semantic dividing method based on deep learning |
| CN111582004A (en) * | 2019-02-15 | 2020-08-25 | 阿里巴巴集团控股有限公司 | Target area segmentation method and device in ground image |
| CN111597861A (en) * | 2019-02-21 | 2020-08-28 | 中科星图股份有限公司 | System and method for automatically interpreting ground object of remote sensing image |
| CN109886238A (en) * | 2019-03-01 | 2019-06-14 | 湖北无垠智探科技发展有限公司 | Unmanned plane Image Change Detection algorithm based on semantic segmentation |
| CN110033045A (en) * | 2019-04-17 | 2019-07-19 | 内蒙古工业大学 | A kind of method and apparatus of trained identification image atomization |
| CN110210513A (en) * | 2019-04-23 | 2019-09-06 | 深圳信息职业技术学院 | Data classification method, device and terminal device |
| CN110335245A (en) * | 2019-05-21 | 2019-10-15 | 青岛科技大学 | Cage netting damage monitoring method and system based on monocular space and time continuous image |
| CN110197505A (en) * | 2019-05-30 | 2019-09-03 | 西安电子科技大学 | Remote sensing images binocular solid matching process based on depth network and semantic information |
| CN110197505B (en) * | 2019-05-30 | 2022-12-02 | 西安电子科技大学 | Remote sensing image binocular stereo matching method based on depth network and semantic information |
| CN110245747A (en) * | 2019-06-21 | 2019-09-17 | 华中师范大学 | Image processing method and device based on full convolutional neural networks |
| CN110245747B (en) * | 2019-06-21 | 2021-10-19 | 华中师范大学 | Image processing method and device based on full convolution neural network |
| CN110532224A (en) * | 2019-08-13 | 2019-12-03 | 武汉中海庭数据技术有限公司 | A kind of file management system and method for deep learning mark sample |
| CN110600105B (en) * | 2019-08-27 | 2022-02-01 | 武汉科技大学 | CT image data processing method, device and storage medium |
| CN110600105A (en) * | 2019-08-27 | 2019-12-20 | 武汉科技大学 | CT image data processing method, device and storage medium |
| CN110765853A (en) * | 2019-09-10 | 2020-02-07 | 中国科学院上海技术物理研究所 | Image processing method of multi-spectrum phase machine |
| CN110765853B (en) * | 2019-09-10 | 2023-05-05 | 中国科学院上海技术物理研究所 | Image processing method of multispectral camera |
| CN112683921A (en) * | 2019-10-17 | 2021-04-20 | 神讯电脑(昆山)有限公司 | Image scanning method and image scanning system for metal surface |
| CN111046835A (en) * | 2019-12-24 | 2020-04-21 | 杭州求是创新健康科技有限公司 | Eyeground illumination multiple disease detection system based on regional feature set neural network |
| CN111242228B (en) * | 2020-01-16 | 2024-02-27 | 武汉轻工大学 | Hyperspectral image classification method, hyperspectral image classification device, hyperspectral image classification equipment and storage medium |
| CN111242228A (en) * | 2020-01-16 | 2020-06-05 | 武汉轻工大学 | Hyperspectral image classification method, device, equipment and storage medium |
| CN111311571A (en) * | 2020-02-13 | 2020-06-19 | 上海小萌科技有限公司 | Target information acquisition method, system, device and readable storage medium |
| CN111505738A (en) * | 2020-03-17 | 2020-08-07 | 上海眼控科技股份有限公司 | Method and equipment for predicting meteorological factors in numerical weather forecast |
| CN111476170A (en) * | 2020-04-09 | 2020-07-31 | 首都师范大学 | Remote sensing image semantic segmentation method combining deep learning and random forest |
| WO2021238992A1 (en) * | 2020-05-26 | 2021-12-02 | 杭州海康威视数字技术股份有限公司 | Neural network training method and apparatus, electronic device, and readable storage medium |
| CN111832569A (en) * | 2020-06-19 | 2020-10-27 | 西北大学 | High-spectrum classification and segmentation based mural paint layer shedding disease labeling method |
| CN111832569B (en) * | 2020-06-19 | 2023-10-27 | 西北大学 | Wall painting pigment layer falling disease labeling method based on hyperspectral classification and segmentation |
| CN113820012A (en) * | 2020-06-19 | 2021-12-21 | 华为技术有限公司 | Multispectral image processing method and device |
| CN111882563A (en) * | 2020-07-13 | 2020-11-03 | 浙江大学 | Semantic segmentation method based on directional convolutional network |
| CN112084923A (en) * | 2020-09-01 | 2020-12-15 | 西安电子科技大学 | Semantic segmentation method for remote sensing image, storage medium and computing device |
| CN112084923B (en) * | 2020-09-01 | 2023-12-22 | 西安电子科技大学 | Remote sensing image semantic segmentation method, storage medium and computing device |
| CN112633185A (en) * | 2020-09-04 | 2021-04-09 | 支付宝(杭州)信息技术有限公司 | Image processing method and device |
| CN111814771A (en) * | 2020-09-04 | 2020-10-23 | 支付宝(杭州)信息技术有限公司 | Image processing method and device |
| CN112348820A (en) * | 2020-11-06 | 2021-02-09 | 天津市勘察设计院集团有限公司 | Remote sensing image semantic segmentation method based on depth discrimination enhancement network |
| CN112348820B (en) * | 2020-11-06 | 2023-04-07 | 天津市勘察设计院集团有限公司 | Remote sensing image semantic segmentation method based on depth discrimination enhancement network |
| CN112784806A (en) * | 2021-02-04 | 2021-05-11 | 中国地质科学院矿产资源研究所 | Lithium-containing pegmatite vein extraction method based on full convolution neural network |
| CN113468951A (en) * | 2021-05-20 | 2021-10-01 | 华东师范大学 | Method for detecting mangrove landform based on Landsat satellite remote sensing |
| CN113327250A (en) * | 2021-05-28 | 2021-08-31 | 深圳前海微众银行股份有限公司 | Multispectral image segmentation method and device, electronic device and storage medium |
| CN113327250B (en) * | 2021-05-28 | 2024-05-31 | 深圳前海微众银行股份有限公司 | Multi-spectral image segmentation method and device, electronic equipment and storage medium |
| CN113627292A (en) * | 2021-07-28 | 2021-11-09 | 广东海启星海洋科技有限公司 | Remote sensing image identification method and device based on converged network |
| CN113627292B (en) * | 2021-07-28 | 2024-04-30 | 广东海启星海洋科技有限公司 | Remote sensing image recognition method and device based on fusion network |
| CN114898097A (en) * | 2022-06-01 | 2022-08-12 | 首都师范大学 | Image recognition method and system |
| CN114898097B (en) * | 2022-06-01 | 2024-05-10 | 首都师范大学 | Image recognition method and system |
| CN115423829A (en) * | 2022-07-29 | 2022-12-02 | 江苏省水利科学研究院 | Method and system for rapidly extracting water body from single-band remote sensing image |
| CN115423829B (en) * | 2022-07-29 | 2024-03-01 | 江苏省水利科学研究院 | Method and system for rapidly extracting water body of single-band remote sensing image |
| CN115661680A (en) * | 2022-11-15 | 2023-01-31 | 北京轨道未来空间科技有限公司 | Satellite remote sensing image processing method |
| CN116883853A (en) * | 2023-01-12 | 2023-10-13 | 河南大学 | Crop space-time information remote sensing classification method based on transfer learning |
| CN116883853B (en) * | 2023-01-12 | 2024-05-28 | 河南大学 | Crop space-time information remote sensing classification method based on transfer learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108564587A (en) | A kind of a wide range of remote sensing image semantic segmentation method based on full convolutional neural networks | |
| Pacifici et al. | An innovative neural-net method to detect temporal changes in high-resolution optical satellite imagery | |
| Samiappan et al. | Mapping of invasive phragmites (common reed) in Gulf of Mexico coastal wetlands using multispectral imagery and small unmanned aerial systems | |
| CN109697475A (en) | A kind of muskeg information analysis method, remote sensing monitoring component and monitoring method | |
| KR100678395B1 (en) | System and method for real time position correction of geometric information data using satellite and aerospace image | |
| CN108195767B (en) | Estuary wetland foreign species monitoring method | |
| CN107392130A (en) | Classification of Multispectral Images method based on threshold adaptive and convolutional neural networks | |
| CN103544505B (en) | Ship identification system and method towards unmanned plane Aerial Images | |
| CN106709517A (en) | Mangrove recognition method and system | |
| CN110060273A (en) | Remote sensing image landslide plotting method based on deep neural network | |
| CN109308451A (en) | A kind of high score data information extraction system and method | |
| CN107688776A (en) | A kind of urban water-body extracting method | |
| Nazmfar et al. | Classification of satellite images in assessing urban land use change using scale optimization in object-oriented processes (a case study: Ardabil city, Iran) | |
| Jiang et al. | A multi-scale approach to detecting standing dead trees in UAV RGB images based on improved faster R-CNN | |
| Rampi et al. | Minnesota Land Cover Classification and Impervious Surface Area by Landsat and Lidar: 2013-14 Update | |
| Cui et al. | Combined Model Color-Correction Method Utilizing External Low-Frequency Reference Signals for Large-Scale Optical Satellite Image Mosaics. | |
| Sakieh et al. | An integrated spectral-textural approach for environmental change monitoring and assessment: analyzing the dynamics of green covers in a highly developing region | |
| CN113627292B (en) | Remote sensing image recognition method and device based on fusion network | |
| Leon et al. | Data fusion for mapping coral reef geomorphic zones: Possibilities and limitations | |
| Jaiswal et al. | Deep learning based windthrow detection for winter storms | |
| Medhi et al. | Rural Road Extraction using Object Based Image Analysis (OBIA): A case study from Assam, India | |
| Ramlal | Utilizing emerging geo-imaging technologies for the management of tropical coastal environments | |
| Ha et al. | A semi‐automatic workflow for plot boundary extraction of irregularly sized and spaced field plots from UAV imagery | |
| CN110033460A (en) | It is a kind of based on scale space transformation satellite image in mariculture area extracting method | |
| Wang et al. | Spartina alterniflora classification at patch scale based on feature fusion and deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180921 |