CN108460403A - The object detection method and system of multi-scale feature fusion in a kind of image - Google Patents
The object detection method and system of multi-scale feature fusion in a kind of image Download PDFInfo
- Publication number
- CN108460403A CN108460403A CN201810065807.1A CN201810065807A CN108460403A CN 108460403 A CN108460403 A CN 108460403A CN 201810065807 A CN201810065807 A CN 201810065807A CN 108460403 A CN108460403 A CN 108460403A
- Authority
- CN
- China
- Prior art keywords
- scale
- target
- image
- fusion
- detection
- 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
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/29—Graphical models, e.g. Bayesian networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/07—Target detection
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses the object detection method and system of multi-scale feature fusion in a kind of image, the method includes：The first step carries out the scaling of different scale using picture to be detected, constructs image pyramid；Second step obtains the multiple scale detecting template of one group of most of sample size of covering using statistics clustering method；Third walks, and is based on multiple scale detecting template, carries out the target context structure of dimension self-adaption；4th step, multiple dimensioned depth characteristic fusion；5th step, the non-maxima suppression based on soft-decision.The present invention is by constructing the serial of methods such as the sparse pyramid of image multiresolution, multiple scale detecting masterplate, masterplate dimension self-adaption context, the fusion of multiple dimensioned depth characteristic, the abundant excavation and fusion for realizing depth characteristic utilize, and can promote target detection performance.
Description
Technical field
The present invention relates to a kind of method of object detection field in image, multiple features fusion in specifically a kind of image
Object detection method and system.
Background technology
Target detection identification in image has extensive functional need in application scenarios such as intelligent video monitorings,
It is also the more popular research direction of computer vision field.Existing image object detection method, because remaining following difficulty
And challenge, testing result are also to be hoisted：(1) between similar target, there are larger more for the appearance features such as color and vein shape
Sample, otherness.(2) similar target causes the structure feature of sample in class to there is large change there are the diversity of posture.Such as
There is target upright, postures, the similar target of different postures such as drop to the ground, tilt will present different profiles, shape in reality
Etc. structure features；(3) similar object height, width equidimension size and ratio constant interval are big.One side target physical height
There will be larger distributed areas, on the other hand due to the different target of shooting distance also will present in the picture it is different big
The dimensional variations such as small, ratio.(4) blocking for target can influence testing result.Its partial information is missing from after target is blocked,
Increase detection difficulty.(5) target local environment background and the diversity of illumination cause flase drop to increase.Target appears in as outdoor
When, such as urban road, entrance, background is often complex, and some complicated background such as trees, street lamp meeting and targets
Generation is obscured, leads to flase drop.
Currently, more mature object detection method can be divided into two classes substantially：(1) it is based on background modeling.This method master
It is used to detect moving target in video：The still image that will be inputted carries out scene cut, utilizes mixed Gauss model (GMM)
Or the methods of motion detection, it is partitioned into its foreground and background, then extract special exercise target in the foreground.Such methods need to connect
Continuous image sequence models to realize, the target detection being not suitable in single image.(2) it is based on statistical learning.Will own
The image collection of known genera Mr. Yu's one kind target gets up to form training set, method (such as HOG, Harr based on an engineer
Deng) to training set image zooming-out feature.The feature of extraction is generally the letters such as gray scale, texture, histogram of gradients, the edge of target
Breath.Then pedestrian detection grader is built according to the feature database of a large amount of training sample.Grader is generally available SVM,
The models such as Adaboost and neural network.
The object detection method performance based on statistical learning is more excellent in recent years in terms of comprehensive, the target inspection based on statistical learning
Survey method can be divided into traditional artificial characteristic target detection method and depth characteristic machine learning object detection method.
Traditional artificial characteristic target detection method is primarily referred to as its feature for utilizing engineer, to carry out target target
Modeling.The characterization method for showing outstanding engineer in recent years includes mainly：Pedro F.Felzenszwalb in 2010 etc.
DPM (Deformable Part Model) method (the Object detection with discriminatively of proposition
trained part-based models).ICF (the Integral Channel proposed in 2009 such as Piotr Doll á r
Features), the ACF methods (Fast Feature Pyramids for Object Detection) proposed in 2014.
Informed Harr methods (the Informed Haar-like Features of the propositions such as Shanshan Zhang in 2014
Improve Pedestrian Detection), the extraction more Harr features with characterization information are dedicated to be trained.
Although the feature of these engineers achieves certain effect, but because manual features characterize scarce capacity, there are still detections
The not high problem of precision.Due to feature learning and ability to express more powerful possessed by depth convolutional neural networks model, scheming
As target classification context of detection obtains more and more extensive and successful application.The target detection operator on basis is R-CNN
(Region-Convolutional Neural Network) model.Girshick in 2014 et al. proposes RCNN for general
The detection of target is again later to propose fast-rcnn and faster-rcnn, improves based on deep learning target detection side
The accuracy and speed of method.The methods of Yolo proposed and SSD in 2016, then realize the fast of single stage by thoughts such as Anchor
Fast target detection.These target detections based on depth learning technology are all based on greatly single scale, fixed size context
Depth characteristic, there are still depth characteristics, and insufficient problem, detection performance to be utilized to need to be further increased.
Invention content
For insufficient existing for the object detection method based on depth model, multiple dimensioned spy in a kind of image of present invention proposition
The object detection method and system for levying fusion, by constructing the sparse pyramid of image multiresolution, multiple scale detecting masterplate, masterplate
A series of creative methods such as dimension self-adaption context, the fusion of multiple dimensioned depth characteristic, realize the abundant excavation of depth characteristic
It is utilized with fusion, promotes target detection performance.
According to the first aspect of the invention, a kind of object detection method of multi-scale feature fusion in image is provided, including：
S1：The scaling of different scale is carried out using picture to be detected, constructs an image pyramid；
S2：Based on the training image that described image pyramid obtains, it is big to obtain one group of covering using statistics clustering method
The multiple scale detecting template of most sample sizes；
S3：On the basis of above-mentioned multiple scale detecting template, the target context structure of dimension self-adaption is carried out；
S4：According to target context structure as a result, carry out multiple dimensioned depth characteristic fusion, Analysis On Multi-scale Features figure is obtained；
S5：According to above-mentioned Analysis On Multi-scale Features figure, the non-maxima suppression based on soft-decision is carried out, is realized multiple dimensioned in image
The target detection of Fusion Features.
Preferably, in the S1：In order to enable detection network utilizes the detection block of one or several limited sizes, it can be right
In image different size of target completely can compactly carry out frame choosing sampling, need to original image carry out it is multiple dimensioned scaling,
So that original object increases its by repeatedly scaling is detected probability of the complete compact frame choosing of frame, by picture to be trained by by than
Example is scaled to the picture of L different resolution size, to the image pyramid of one resolution ratio of construction from high to low.Specifically
, when training, for each original training image, the scaling of multiple scales is carried out, obtains L under different scale
Image is for training.When test, for each image to be detected, the scaling of multiple scales is equally carried out, is obtained not
With L image under scale for detecting, and to the testing result of this L image, amalgamation judging is carried out, final detection is obtained
As a result.
Preferably, the multiple scale detecting mould that one group of most of sample size of covering is obtained using statistics clustering method
Plate refers to：Based on K-medoids clustering methods, and using Jie Kade distances (Jaccard distance) as Cluster Assessment
Index concentrates target to be clustered by different wide high level and the ratio of width to height training data, forms the width of one group of K cluster centre
High ratio, the target template as the most wide high proportions of covering.
Preferably, the target context for carrying out dimension self-adaption is built, and refers to：By the output of CNN network convolutional layers
The receptive field each put on characteristic pattern, as candidate target frame；The part that receptive field has more relative to pattern plate bolster is used as target
The context of frame, the detection for being used for auxiliary mark identify.
It is highly preferred that it is described carry out dimension self-adaption target context structure, finally obtain a contextual information with
Target scale changes and the detection model of variation, i.e.,：Small scaled target will obtain the contextual information of bigger, and large scale target
Contextual information it is less, to meet different demands of the target to contextual information of different scale.
Preferably, described to carry out multiple dimensioned depth characteristic fusion, refer to：The characteristic pattern that CNN difference convolutional layers are exported, choosing
Go out M layers to be merged, for constructing Analysis On Multi-scale Features pyramid.
It is highly preferred that described carry out multiple dimensioned depth characteristic fusion, specially：It is in the convolutional layer selected for M
Last layer of CNN networks, the characteristic pattern output it up-sample it using deconvolution, it is made to be expanded to and last layer feature
It after scheming same resolution sizes, does with last layer characteristic pattern and is added pixel-by-pixel, obtain the Analysis On Multi-scale Features figure of fusion adjacent two layers；
Again and so on, deconvolution expands, is merged with more last layer characteristic pattern, until the fusion for all M layers of characteristic pattern for completing to select.
Preferably, the non-maxima suppression based on soft-decision refers to：
The maximum detection block of fiducial probability is first elected, other detection blocks and the maximum detection block meter of fiducial probability are passed through
IOU (intersection over union) is calculated, then reduces its fiducial probability more than a certain threshold value；
After the maximum detection block of this fiducial probability is removed, then select the maximum detection of fiducial probability in remaining detection block
Frame, and remaining detection block and the maximum detection block of fiducial probability are calculated into IOU, then its fiducial probability is dropped more than a certain threshold value
It is low；
By the above continuous iteration, the detection block after screening to the end is obtained.
According to the second aspect of the invention, a kind of object detection system of multi-scale feature fusion in image is provided, including：
Image pyramid builds module：The scaling of different scale is carried out using picture to be detected, constructs an image gold word
Tower；
Multiple scale detecting template builds module：The instruction that the described image pyramid of module obtains is built based on image pyramid
Practice image, the multiple scale detecting template of one group of most of sample size of covering is obtained using statistics clustering method；
Target context builds module：In the multiple scale detecting template that above-mentioned multiple scale detecting template structure module obtains
On the basis of, carry out the target context structure of dimension self-adaption；
Multiple dimensioned depth characteristic melts module：According to target context structure module as a result, carrying out multiple dimensioned depth characteristic
Fusion, obtains Analysis On Multi-scale Features figure；
Module of target detection：The Analysis On Multi-scale Features figure for melting module according to above-mentioned multiple dimensioned depth characteristic sentenced based on soft
Non-maxima suppression certainly realizes the target detection of multi-scale feature fusion in image.
Preferably, described image pyramid construction module, when training, for each original training image, into
The scaling of the multiple scales of row obtains L image under different scale for training.When test, for each to be detected
Image, equally carry out the scaling of multiple scales, obtain L image under different scale for detecting, and to this L image
Testing result carries out amalgamation judging, obtains final testing result.
Preferably, the multiple scale detecting template builds module, which is based on K-medoids clustering methods, and utilizes
Jie Kade distances (Jaccard distance) are used as Cluster Assessment index, concentrate target to press different wide high level training data
And the ratio of width to height is clustered, and the ratio of width to height of one group of K cluster centre is formed, the target mould as the most wide high proportions of covering
Plate.
Preferably, the target context builds module, on the characteristic pattern which export CNN network convolutional layers each
The receptive field of point, as candidate target frame；The part that receptive field has more relative to pattern plate bolster, the i.e. context as target frame,
It is identified for the detection of auxiliary mark.
Preferably, the multiple dimensioned depth characteristic melts module, the characteristic pattern which exports CNN difference convolutional layers, choosing
Go out M layers to be merged, for constructing Analysis On Multi-scale Features pyramid.
It is highly preferred that the multiple dimensioned depth characteristic melts module, CNN is in the convolutional layer which selects M
Last layer of network, the characteristic pattern output it up-sample it using deconvolution, it is made to be expanded to and last layer characteristic pattern
It after same resolution sizes, does with last layer characteristic pattern and is added pixel-by-pixel, obtain the Analysis On Multi-scale Features figure of fusion adjacent two layers；Again
And so on, deconvolution expands, is merged with more last layer characteristic pattern, until the fusion for all M layers of characteristic pattern for completing to select.
Preferably, the module of target detection first elects the maximum detection block of fiducial probability, is detected by others
Frame calculates IOU (intersection over union) with the maximum detection block of fiducial probability, is then set more than a certain threshold value
Believe that probability reduces；
After the maximum detection block of this fiducial probability is removed, then select the maximum detection of fiducial probability in remaining detection block
Frame, and remaining detection block and the maximum detection block of fiducial probability are calculated into IOU, then its fiducial probability is dropped more than a certain threshold value
It is low；
By the above continuous iteration, the detection block after screening to the end is obtained.
Compared with prior art, the invention has the advantages that：
According to the target identification of different scale to the difference of contextual information demand in the present invention, the structure in conjunction with CNN is special
The concept of point and receptive field is modeled using the contextual information of dimension self-adaption, is identified for auxiliary mark.
In the present invention in such a way that K-medoids clustering methods optimize the selection of template, meets different templates and target is examined
Survey model inspection effect.
Image pyramid solves the problems, such as that training picture needs fixed size, design to be obtained using the method for crop in the present invention
Picture to fixed size is used to train.
Adequately utilize the feature of CNN different layers can be simultaneously by further feature and shallow-layer Fusion Features in the present invention
Using the detailed information of the characterization ability and shallow-layer feature of further feature, the accuracy of detection of small scaled target is promoted.
Analysis On Multi-scale Features testing mechanism is established in the present invention, will again be merged after feature combination, is then examined again
It surveys, finally merges the result of detection, realize Analysis On Multi-scale Features detection.
The non-maxima suppression based on soft-decision is utilized in the present invention, improves syncretizing effect.
To sum up, the present invention is by comprehensively utilizing multi-resolution pyramid, multiple dimensioned template cluster, dimension self-adaption context
The technologies such as information, the fusion of multiple dimensioned depth characteristic, soft-decision non-maxima suppression, enhance feature learning and the expression of image object
Ability, effectively promotes the precision of the target detections such as pedestrian, at the same preferably solve in the prior art target in small scale, long distance
From when, the problems such as intensive target detection.
Description of the drawings
Upon reading the detailed description of non-limiting embodiments with reference to the following drawings, other feature of the invention,
Objects and advantages will become more apparent upon：
Fig. 1 a, Fig. 1 b are that image pyramid builds flow chart in one embodiment of the invention；
Fig. 2 is that multiple dimensioned template obtains flow chart in one embodiment of the invention；
Fig. 3 is multi-scale feature fusion implementation flow chart in one embodiment of the invention；
Fig. 4 is the non-maxima suppression implementation flow chart of soft-decision in one embodiment of the invention.
Specific implementation mode
With reference to specific embodiment, the present invention is described in detail.Following embodiment will be helpful to the technology of this field
Personnel further understand the present invention, but the invention is not limited in any way.It should be pointed out that the ordinary skill of this field
For personnel, without departing from the inventive concept of the premise, various modifications and improvements can be made.These belong to the present invention
Protection domain.
Existing object detection method can identify certain larger-size targets well, but size is larger
Target small part is only accounted in actual life, for distance target farther out, testing result is not fine.Target
Detection has following characteristics, by taking pedestrian target as an example：
Feature one, the diversity of scale.Have old man, a middle-aged person, child in one side pedestrian, physics height there will be compared with
Big distributed area.On the other hand due to the difference of shooting distance, for the same pedestrian, higher, the remoter bat of distance
It takes the photograph and a little takes that pedestrian its pixel is lower, and height is lower；And pedestrian's pixel that lower, closer shooting point takes is higher, it is high
Degree is closer to its true height.The pedestrian detection effect that existing method is more than height 100 pixels is preferable, but for remote
Distance, the low pedestrian detection effect of height are bad.All it is often desirable when pedestrian detection is applied in vehicle DAS (Driver Assistant System)
The pedestrian that system detectio distant location occurs reminds driver, therefore solves the problems, such as that the remote pedestrian of detection is also one urgent
The demand cut.
Feature two is blocked.The pedestrian taken in reality can have circumstance of occlusion, on the one hand can there is the feelings that crowd clusters round
Condition.When multiple people walk together, look over always that someone will be blocked a part for body from all angles；Another aspect pedestrian
May a part for body be sheltered from by the object in environment, such as number, vehicle, house.Its body after pedestrian is blocked at this time
Body portion information is missing from, and the result of missing inspection can be then caused for the detector based on human body integrity profile feature.
The present invention proposes that object detection method based on multi-feature fusion is preferably solved from the above problem in image
Target of having determined small scale, it is remote when, the problems such as intensive target detection.It is tired that the present invention is based on target detections present in reality
It is difficult, it is proposed that it is merged for multiscale target self-adapting detecting, and using multilayer receptive field difference based on pyramid, and
Using soft-decision non-maxima suppression for intensive target detection and then the effect of promotion target detection.Wherein：
The first step, the scaling of different scale is carried out using picture to be detected, and construction includes the figure from the multiple resolution ratio of size
As pyramid；
Second step clusters sample according to the wide high size of training sample, passes through cluster using statistics clustering method
Form K cluster centre, each cluster centre forms one using high mean value not lend oneself to worry and anxiety in this as the detection template of scale；K cluster
The template group for forming one group of K different scale, as detector target template group；
Third walks, the target context structure of dimension self-adaption.
To the detection template frame of each scale, all vertically left and right directions uniform expansions expands to same convolutional neural networks
The receptive field size that the characteristic pattern that (CNN, Convolutional Neural Network) is finally exported each is put is identical, extension
Part forms the contextual information for including target, and the context size for including is adaptive with template size.
4th step, multi-scale feature fusion.
By the way that the more a convolutional layer characteristic patterns of CNN are added fusion pixel-by-pixel in same resolution ratio, target inspection is formed
Survey the multiple dimensioned depth characteristic needed for identification.
5th step, the non-maxima suppression based on soft-decision.
Confidence level by reducing detection block replaces directly deleting detection block, then is sieved to detection block by continuous iteration
Choosing.
Meanwhile the present invention constructs an image object detecting system, system synthesis profit by integrated above method step
With multi-resolution pyramid, multiple dimensioned template cluster, the fusion of dimension self-adaption contextual information, multiple dimensioned depth characteristic, soft sentence
Certainly the methods of non-maxima suppression enhances the feature learning and expression ability of image object, effectively promotes the target detections such as pedestrian
Precision.
Specifically, object detection system based on multi-feature fusion in a kind of image, including：
Image pyramid builds module：The scaling of different scale is carried out using picture to be detected, constructs an image gold word
Tower；
Multiple scale detecting template builds module：The instruction that the described image pyramid of module obtains is built based on image pyramid
Practice image, the multiple scale detecting template of one group of most of sample size of covering is obtained using statistics clustering method；
Target context builds module：In the multiple scale detecting template that above-mentioned multiple scale detecting template structure module obtains
On the basis of, carry out the target context structure of dimension self-adaption；
Multiple dimensioned depth characteristic melts module：According to target context structure module as a result, carrying out multiple dimensioned depth characteristic
Fusion, obtains Analysis On Multi-scale Features figure；
Module of target detection：The Analysis On Multi-scale Features figure for melting module according to above-mentioned multiple dimensioned depth characteristic sentenced based on soft
Non-maxima suppression certainly realizes the target detection of multi-scale feature fusion in image.
The above method of the present invention and system are described in detail by taking pedestrian detection as an example below, especially above-mentioned five
Implementation involved in a step/module.
One, the scaling that different scale is carried out using picture to be detected constructs image pyramid.
In order to enable detection network utilizes the detection block of one or several limited sizes, it can be to different size of in image
Target completely can compactly carry out frame choosing sampling, need to carry out multiple dimensioned scaling, so that original object is passed through to original image
Repeatedly scaling increases its detected frame probability that completely compact frame selects.Will picture be trained by be scaled out into L difference
The picture of resolution sizes, to the image pyramid of one resolution ratio of construction from high to low.
When the training of pedestrian's detection model and test, initial data is zoomed in and out using the principle of image pyramid
Processing, as shown in attached drawing 1a, 1b.
It is exactly specifically, for each image to be trained, to carry out the contracting of 0.5X, 1X and 2X when training
It puts, obtains the image under different scale for training, as shown in fig. la.When test, for each figure to be detected
Picture carries out the scaling of 0.5X, 1X and 2X, obtains the image under different scale for detecting, and to the detection of three scale images
As a result, carrying out amalgamation judging, final testing result is obtained, as shown in figure ib.
In the present embodiment, the gradient of backpropagation is calculated for convenience, so the picture of input needs fixed scale,
Such as the size of all pictures is all 640*480 pixels in Caltech.After random scaling, training picture scale has occurred
The picture of variation, 640*480 sizes becomes 320*240 after 0.5X scalings, and scale then becomes after passing through 2X scalings
For 1280*960.In order to obtain the training picture of unified scale, when training, ruler is cut out from the picture by scaling
Degree is the picture of 640*480, and concrete operations flow is as shown in Figure 1a.I.e. for the picture that size is 320*240, image completion is used
To 640*480, and for the picture that size is 1280*960, then therefrom random cropping goes out the small figure of 640*480, then by this three
Kind picture is simultaneously for training, as shown in fig. la.This way can be effectively increased training samples number, promote deep learning etc.
The performance of data-driven method.
Two, best detector template group is obtained using statistics clustering method.
The present invention obtains best detector template group using statistics clustering method, has been demarcated in extraction training image
Frame progress K-medoids clustering methods are obtained not same group of (K) multiple dimensioned template by target rectangle frame.Here scale is
The size (long and wide) and ratio (the ratio of width to height) of feeling the pulse with the finger-tip mark frame.K multiple dimensioned templates are obtained using statistical method, had made it both
The differentiation that most training sample scales take into account sample size in class again can be covered.To consider when choosing template
The differentiation of sample size, avoids only being caused that matching template is very few, is difficult to meet multiple dimensioned mesh with single or a small number of scale templates
Mark and the accurate match cognization problem of template, but statistical data concentrates wide high and ratio the distribution situation of target sample, by poly-
Class forms multiple cluster centres, each cluster centre forms a scale (average the ratio of width to height but not limited to this) template, to root
One group of multiple dimensioned template is formed according to sample size distribution；Meanwhile the quantity by limiting cluster centre, to avoid center excessive
Lead to that the template training sample of a certain scale is very little, detector cannot train up problem.K- can be used in clustering method
Medoids etc. is realized and is carried out cluster and the selection of multiple dimensioned template to wide high and the ratio of width to height of sample that data are concentrated.
When specific operation, can first definition template scale, cover the pedestrian target of different scale.Generally vertical pedestrian's
The ratio of width to height is generally 1:3, according to this experience and the general height of pedestrian target in Caltech data sets is combined to be distributed, it can be with
It manually goes to select typical template scale, such as 30*90,50*150, but this mode chosen by hand does not have not only
Versatility, and very maximum probability cannot choose most suitable template.This is because：
First, in different data sets, due to the difference of photo resolution and monitoring visual angle, the distribution of pedestrian level is
Different, this method for choosing template by hand does not have versatility；
Secondly, in true monitoring scene, because of situations such as the appearance of pedestrian's posture, camera angle, block information,
Cause the ratio of width to height of pedestrian to vary widely, the template chosen is caused not have typicalness.Meanwhile when choosing template
Consider the distribution situation of object height in data set, if the corresponding training sample of the template of a certain scale is very little, can cause
The corresponding detector of this template cannot be trained adequately.
Therefore, propose that the method clustered by K-medoids is high to the width of the sample in pedestrian's data set in the present embodiment
Clustered, using statistical method obtain the multiple dimensioned templates of K (the present embodiment K=32, certainly, in other embodiments
Can also be other numbers).
As shown in Fig. 2, being based on K-medoids clustering methods, pedestrian level and width is concentrated to cluster training data.
Using Jie Kade distances (Jaccard distance) as the Cluster Assessment index of K-medoids, i.e.,：
d(si,sj)=1-J (si,sj)
Wherein, si=(hi,wi)；sj=(hj,wj) indicating two different pedestrian's frames in data set, h, w indicate capable respectively
The height and the width of people's frame, J indicate the Jie Kade similarities (Jaccard similarity coefficient) of standard,
Three, the target context structure of dimension self-adaption.
The present invention using dimension self-adaption target context build, on the one hand, small scaled target because useful information compared with
Less, more contextual informations are generally required to assist in identifying；And large scale target does not often need a large amount of context then.Separately
On the one hand, since each characteristic point of the full articulamentums of CNN corresponds to a fixed receptive field of scale in artwork, receptive field can be based on
To construct the context of different templates.
Specifically, the pattern plate bolster of the K different scale returned out to K-medoids clustering procedures, vertically controls and is expanded
It opens up and to expand to the receptive field each put with the full articulamentums of CNN in the same size.Because receptive field size is fixed, then small scale
Pattern plate bolster need to do the extension of large scale, thus the context of large scale can be obtained；And large scale template is then on the contrary, obtain
The context obtained is smaller.Based on this method, the contextual information with pattern plate bolster dimension self-adaption is can get, is known for auxiliary mark
Not.
Four, multiple dimensioned depth characteristic fusion.
In view of the characteristic pattern of CNN each convolutional layer output, all containing the useful feature on different scale, shallow-layer it is defeated
Go out to usually contain more local detail features, and high-rise feature usually contains more of overall importance and Semantic information.It is right
The feature for the different scale that these different layers obtain carries out fusion utilization and can be obtained more rich feature representation.Based on the think of
Think that the present invention devises a kind of Multiscale Fusion method：
As shown in figure 3, M main convolutional layer (often rearmost M layer in CNN networks, this implementation for CNN
Example is the characteristic pattern of 3) output, shallow from being deep to, and first carries out deconvolution up-sampling to the characteristic pattern of bottommost layer each with deconvolution,
So that it is transformed in the same last layer output same resolution-scale of characteristic pattern, then does the characteristic pattern of this two layers same scale
It is added pixel-by-pixel, the Analysis On Multi-scale Features figure after being merged；Analogize accordingly again, realize all M layers of fusions, obtains M layers of fusion
The Analysis On Multi-scale Features figure of feature.Such as input picture passes through after Resnet networks, by res5 layers of output result deconvolution one
It is secondary to obtain resolution ratio identical with res4, and with the results added of res4, then its deconvolution once obtained identical with res3
Resolution ratio simultaneously obtains final result and is tested with the results added of res3 progress Fusion Features.
Five, the non-maxima suppression of soft-decision.
In pervious method, final testing result is obtained with non-maxima suppression method fusion detection frame.Traditional
Non-maxima suppression method be it is a kind of volume method is sentenced based on the hard of greed, correct detection is may result in during fusion
Frame is suppressed, especially when IOU (Intersection of Units, the overlapping region between two rectangle frames and combined region
Ratio) threshold value selection it is inappropriate when.
As shown in figure 4, the non-maxima suppression Integral Thought of soft-decision is exactly in the present invention：By reducing setting for detection block
Reliability replaces directly deleting detection block.
In the present embodiment, concrete operations are：The maximum detection block of fiducial probability is first elected, is then detected by others
The frame of frame and maximum fiducial probability calculates IOU, then reduces its fiducial probability more than a certain threshold value.Then this maximum frame is removed
The maximum fiducial probability frame in selecting remaining detection block afterwards, and remaining detection block and the frame of maximum fiducial probability are calculated into IOU,
Then its fiducial probability is reduced more than a certain threshold value, by continuous iteration, obtains the detection block after screening to the end.
The mode that linear weighted function may be used in the present embodiment reduces the score of detection block according to the value of iou：
In the formula, M indicates the maximum detection block of fiducial probability, biIndicate i-th of detection block, NiIndicate that IOU threshold values (can
For empirical value or preset value), iou (M, bi) indicate M and biIOU values.
Therefore, soft-decision of the present invention is more preferable for dense population detection result.
Some details of the invention implemented and preferred feature have been carried out in detail above by for pedestrian detection
Thin description.The present invention can also be applied to target detection in other images, however it is not limited to pedestrian detection, the behaviour of other target detections
Make similar to the above embodiments, no longer separately illustrates herein.
To sum up, the present invention is by constructing the sparse pyramid of image multiresolution, the multiple scale detecting mould based on cluster optimization
A series of creative methods such as version, the fusion of masterplate dimension self-adaption context, multiple dimensioned depth characteristic, realize filling for depth characteristic
Divide to excavate and merge and utilize, obtains the promotion of target detection performance.
Specific embodiments of the present invention are described above.It is to be appreciated that the invention is not limited in above-mentioned
Particular implementation, those skilled in the art can make various deformations or amendments within the scope of the claims, this not shadow
Ring the substantive content of the present invention.
Claims (14)
1. the object detection method of multi-scale feature fusion in a kind of image, which is characterized in that including：
S1：The scaling of different scale is carried out using picture to be detected, constructs an image pyramid；
S2：Based on the training image that described image pyramid obtains, most of one group of covering is obtained using statistics clustering method
The multiple scale detecting template of sample size；
S3：On the basis of above-mentioned multiple scale detecting template, the target context structure of dimension self-adaption is carried out；
S4：According to target context structure as a result, carry out multiple dimensioned depth characteristic fusion, Analysis On Multi-scale Features figure is obtained；
S5：According to above-mentioned Analysis On Multi-scale Features figure, the non-maxima suppression based on soft-decision is carried out, realizes Analysis On Multi-scale Features in image
The target detection of fusion.
2. the object detection method of multi-scale feature fusion in image according to claim 1, which is characterized in that the S1
In：In order to enable detection network utilizes the detection block of one or several limited sizes, it can be to different size of target in image
Frame choosing sampling completely can be compactly carried out, needs to carry out multiple dimensioned scaling to original image, make original object by multiple
Scaling increases it and is detected the frame probability that completely compact frame selects, by original image by being scaled out into L different resolution
The picture of size, to the image pyramid of one resolution ratio of construction from high to low.
3. the object detection method of multi-scale feature fusion in image according to claim 1, which is characterized in that the profit
With the multiple scale detecting template of statistics clustering method one group of most of sample size of covering of acquisition, refer to：Based on K-medoids
Clustering method, and Cluster Assessment index is used as using Jie Kade distances, to training data concentrate target press different wide high level and
The ratio of width to height is clustered, and the ratio of width to height of one group of K cluster centre is formed, the target mould as the most wide high proportions of covering
Plate.
4. the object detection method of multi-scale feature fusion in image according to claim 1, which is characterized in that it is described into
The target context of row dimension self-adaption is built, and refers to：The impression that will each be put on the characteristic pattern of CNN network convolutional layers output
Open country, as candidate target frame；The part that receptive field has more relative to pattern plate bolster, the i.e. context as target frame, for assisting
The detection of target identifies.
5. the object detection method of multi-scale feature fusion in image according to claim 4, which is characterized in that it is described into
The target context of row dimension self-adaption is built, and finally obtains the detection that a contextual information changes with target scale variation
Model, i.e.,：Small scaled target will obtain the contextual information of bigger, and the contextual information of large scale target is less, to full
Different demands of the target of sufficient different scale to contextual information.
6. the object detection method of multi-scale feature fusion in image according to claim 1, which is characterized in that it is described into
The multiple dimensioned depth characteristic of row merges, and refers to：The characteristic pattern that CNN difference convolutional layers are exported, selects M layers and is merged, and is used for structure
Make Analysis On Multi-scale Features pyramid.
7. the object detection method of multi-scale feature fusion in image according to claim 6, which is characterized in that it is described into
The multiple dimensioned depth characteristic fusion of row, specially：In last layer of CNN networks in the convolutional layer selected for M, output it
Characteristic pattern it is up-sampled using deconvolution, it is and upper after so that it is expanded to resolution sizes same as last layer characteristic pattern
One layer of characteristic pattern does and is added pixel-by-pixel, obtains the Analysis On Multi-scale Features figure of fusion adjacent two layers；Again and so on, deconvolution expansion,
It is merged with more last layer characteristic pattern, until the fusion for all M layers of characteristic pattern for completing to select.
8. according to the object detection method of multi-scale feature fusion in claim 1-7 any one of them images, feature exists
In the non-maxima suppression based on soft-decision refers to：
The maximum detection block of fiducial probability is first elected, is calculated by other detection blocks and the maximum detection block of fiducial probability
IOU then reduces its fiducial probability more than a certain threshold value；
After the maximum detection block of this fiducial probability is removed, then the maximum detection block of fiducial probability in remaining detection block is selected, and
Remaining detection block and the maximum detection block of fiducial probability are calculated into IOU, then reduce its fiducial probability more than a certain threshold value；
By the above continuous iteration, the detection block after screening to the end is obtained.
9. the object detection system of multi-scale feature fusion in a kind of image, which is characterized in that including：
Image pyramid builds module：The scaling of different scale is carried out using picture to be detected, one image pyramid of construction is more
Size measurement template builds module：The training image that the described image pyramid of module obtains is built based on image pyramid, profit
The multiple scale detecting template of one group of most of sample size of covering is obtained with statistics clustering method；
Target context builds module：On the basis for the multiple scale detecting template that above-mentioned multiple scale detecting template structure module obtains
On, carry out the target context structure of dimension self-adaption；
Multiple dimensioned depth characteristic melts module：According to target context build module as a result, carry out multiple dimensioned depth characteristic fusion,
Obtain Analysis On Multi-scale Features figure；
Module of target detection：The Analysis On Multi-scale Features figure for melting module according to above-mentioned multiple dimensioned depth characteristic, carries out based on soft-decision
Non-maxima suppression realizes the target detection of multi-scale feature fusion in image.
10. the object detection system of multi-scale feature fusion in a kind of image according to claim 9, which is characterized in that
The multiple scale detecting template builds module, which is based on K-medoids clustering methods, and using Jie Kade distances as poly-
Class evaluation index concentrates target to be clustered by different wide high level and the ratio of width to height training data, is formed in one group of K cluster
The ratio of width to height of the heart, the target template as the most wide high proportions of covering.
11. the object detection system of multi-scale feature fusion in a kind of image according to claim 9, which is characterized in that
The target context builds module, the receptive field which will each put on characteristic pattern that CNN network convolutional layers export, as
Candidate target frame；The part that receptive field has more relative to pattern plate bolster, the i.e. context as target frame are used for the inspection of auxiliary mark
Survey identification.
12. the object detection system of multi-scale feature fusion in a kind of image according to claim 9, which is characterized in that
The multiple dimensioned depth characteristic melts module, and the characteristic pattern which exports CNN difference convolutional layers is selected M layers and merged, and uses
In construction Analysis On Multi-scale Features pyramid.
13. the object detection method of multi-scale feature fusion in image according to claim 12, which is characterized in that described
Multiple dimensioned depth characteristic melts module, in last layer of CNN networks in the convolutional layer which selects M, outputs it
Characteristic pattern it is up-sampled using deconvolution, it is and upper after so that it is expanded to resolution sizes same as last layer characteristic pattern
One layer of characteristic pattern does and is added pixel-by-pixel, obtains the Analysis On Multi-scale Features figure of fusion adjacent two layers；Again and so on, deconvolution expansion,
It is merged with more last layer characteristic pattern, until the fusion for all M layers of characteristic pattern for completing to select.
14. according to the object detection system of multi-scale feature fusion in a kind of image of claim 9-13 any one of them,
Be characterized in that, the module of target detection, first elect the maximum detection block of fiducial probability, by other detection blocks with set
Believe that the detection block of maximum probability calculates IOU, then reduces its fiducial probability more than a certain threshold value；
After the maximum detection block of this fiducial probability is removed, then the maximum detection block of fiducial probability in remaining detection block is selected, and
Remaining detection block and the maximum detection block of fiducial probability are calculated into IOU, then reduce its fiducial probability more than a certain threshold value；
By the above continuous iteration, the detection block after screening to the end is obtained.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810065807.1A CN108460403A (en) | 2018-01-23 | 2018-01-23 | The object detection method and system of multi-scale feature fusion in a kind of image |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810065807.1A CN108460403A (en) | 2018-01-23 | 2018-01-23 | The object detection method and system of multi-scale feature fusion in a kind of image |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108460403A true CN108460403A (en) | 2018-08-28 |
Family
ID=63238616
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810065807.1A Pending CN108460403A (en) | 2018-01-23 | 2018-01-23 | The object detection method and system of multi-scale feature fusion in a kind of image |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108460403A (en) |
Cited By (51)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109241902A (en) * | 2018-08-30 | 2019-01-18 | 北京航空航天大学 | A kind of landslide detection method based on multi-scale feature fusion |
| CN109255352A (en) * | 2018-09-07 | 2019-01-22 | 北京旷视科技有限公司 | Object detection method, apparatus and system |
| CN109255320A (en) * | 2018-09-03 | 2019-01-22 | 电子科技大学 | A kind of improved non-maxima suppression method |
| CN109308458A (en) * | 2018-08-31 | 2019-02-05 | 电子科技大学 | A method of small target deteection precision is promoted based on characteristic spectrum change of scale |
| CN109343701A (en) * | 2018-09-03 | 2019-02-15 | 电子科技大学 | A kind of intelligent human-machine interaction method based on dynamic hand gesture recognition |
| CN109363614A (en) * | 2018-08-29 | 2019-02-22 | 合肥德易电子有限公司 | Intelligent integral robot cavity mirror system with high definition video enhancing processing function |
| CN109472193A (en) * | 2018-09-21 | 2019-03-15 | 北京飞搜科技有限公司 | Method for detecting human face and device |
| CN109522958A (en) * | 2018-11-16 | 2019-03-26 | 中山大学 | Based on the depth convolutional neural networks object detection method merged across scale feature |
| CN109598290A (en) * | 2018-11-22 | 2019-04-09 | 上海交通大学 | A kind of image small target detecting method combined based on hierarchical detection |
| CN109615016A (en) * | 2018-12-20 | 2019-04-12 | 北京理工大学 | A kind of object detection method of the convolutional neural networks based on pyramid input gain |
| CN109635717A (en) * | 2018-12-10 | 2019-04-16 | 天津工业大学 | A kind of mining pedestrian detection method based on deep learning |
| CN109658412A (en) * | 2018-11-30 | 2019-04-19 | 湖南视比特机器人有限公司 | It is a kind of towards de-stacking sorting packing case quickly identify dividing method |
| CN109711427A (en) * | 2018-11-19 | 2019-05-03 | 深圳市华尊科技股份有限公司 | Object detection method and Related product |
| CN109815868A (en) * | 2019-01-15 | 2019-05-28 | 腾讯科技（深圳）有限公司 | A kind of image object detection method, device and storage medium |
| CN109816671A (en) * | 2019-01-31 | 2019-05-28 | 深兰科技（上海）有限公司 | A kind of object detection method, device and storage medium |
| CN109816036A (en) * | 2019-01-31 | 2019-05-28 | 北京字节跳动网络技术有限公司 | Image processing method and device |
| CN109886147A (en) * | 2019-01-29 | 2019-06-14 | 电子科技大学 | A kind of more attribute detection methods of vehicle based on the study of single network multiple-task |
| CN109978035A (en) * | 2019-03-18 | 2019-07-05 | 西安电子科技大学 | Pedestrian detection method based on improved k-means and loss function |
| CN110046550A (en) * | 2019-03-14 | 2019-07-23 | 中山大学 | Pedestrian's Attribute Recognition system and method based on multilayer feature study |
| CN110110599A (en) * | 2019-04-03 | 2019-08-09 | 天津大学 | A kind of Remote Sensing Target detection method based on multi-scale feature fusion |
| CN110189307A (en) * | 2019-05-14 | 2019-08-30 | 慧影医疗科技（北京）有限公司 | A kind of pulmonary nodule detection method and system based on multi-model fusion |
| CN110321864A (en) * | 2019-07-09 | 2019-10-11 | 西北工业大学 | Remote sensing images explanatory note generation method based on multiple dimensioned cutting mechanism |
| CN110503098A (en) * | 2019-08-29 | 2019-11-26 | 西安电子科技大学 | A kind of object detection method and equipment of quick real-time lightweight |
| CN110738208A (en) * | 2019-10-08 | 2020-01-31 | 创新奇智(重庆)科技有限公司 | efficient scale-normalized target detection training method |
| CN110796039A (en) * | 2019-10-15 | 2020-02-14 | 北京达佳互联信息技术有限公司 | Face flaw detection method and device, electronic equipment and storage medium |
| CN110796360A (en) * | 2019-10-24 | 2020-02-14 | 吉林化工学院 | Fixed traffic detection source multi-scale data fusion method |
| CN110909615A (en) * | 2019-10-28 | 2020-03-24 | 西安交通大学 | Target detection method based on multi-scale input mixed perception neural network |
| CN110991359A (en) * | 2019-12-06 | 2020-04-10 | 重庆市地理信息和遥感应用中心(重庆市测绘产品质量检验测试中心) | Satellite image target detection method based on multi-scale depth convolution neural network |
| CN111047655A (en) * | 2020-01-10 | 2020-04-21 | 北京盛开互动科技有限公司 | High-definition camera cloth defect detection method based on convolutional neural network |
| CN111127458A (en) * | 2019-12-27 | 2020-05-08 | 深圳力维智联技术有限公司 | Target detection method and device based on image pyramid and storage medium |
| CN111144248A (en) * | 2019-12-16 | 2020-05-12 | 上海交通大学 | People counting method, system and medium based on ST-FHCD network model |
| CN111191621A (en) * | 2020-01-03 | 2020-05-22 | 北京同方软件有限公司 | Rapid and accurate identification method for multi-scale target under large-focus monitoring scene |
| CN111241893A (en) * | 2018-11-29 | 2020-06-05 | 阿里巴巴集团控股有限公司 | Identification recognition method, device and system |
| CN111260615A (en) * | 2020-01-13 | 2020-06-09 | 重庆交通大学 | Laser and machine vision fusion-based method for detecting apparent diseases of unmanned aerial vehicle bridge |
| CN111368600A (en) * | 2018-12-26 | 2020-07-03 | 北京眼神智能科技有限公司 | Method and device for detecting and identifying remote sensing image target, readable storage medium and equipment |
| CN111582012A (en) * | 2019-12-24 | 2020-08-25 | 珠海大横琴科技发展有限公司 | Method and device for detecting small target ship |
| CN111738036A (en) * | 2019-03-25 | 2020-10-02 | 北京四维图新科技股份有限公司 | Image processing method, device, equipment and storage medium |
| CN111767934A (en) * | 2019-10-31 | 2020-10-13 | 杭州海康威视数字技术股份有限公司 | Image identification method and device and electronic equipment |
| CN111986255A (en) * | 2020-09-07 | 2020-11-24 | 北京凌云光技术集团有限责任公司 | Multi-scale anchor initialization method and device of image detection model |
| CN112418208A (en) * | 2020-12-11 | 2021-02-26 | 华中科技大学 | Tiny-YOLO v 3-based weld film character recognition method |
| CN112528976A (en) * | 2021-02-09 | 2021-03-19 | 北京世纪好未来教育科技有限公司 | Text detection model generation method and text detection method |
| CN112560732A (en) * | 2020-12-22 | 2021-03-26 | 电子科技大学中山学院 | Multi-scale feature extraction network and feature extraction method thereof |
| CN112580721A (en) * | 2020-12-19 | 2021-03-30 | 北京联合大学 | Target key point detection method based on multi-resolution feature fusion |
| CN112906677A (en) * | 2021-05-06 | 2021-06-04 | 南京信息工程大学 | Pedestrian target detection and re-identification method based on improved SSD (solid State disk) network |
| CN113222916A (en) * | 2021-04-28 | 2021-08-06 | 北京百度网讯科技有限公司 | Method, apparatus, device and medium for detecting image using target detection model |
| CN113221752A (en) * | 2021-05-13 | 2021-08-06 | 北京惠朗时代科技有限公司 | Multi-template matching-based multi-scale character accurate identification method |
| CN113313213A (en) * | 2021-07-28 | 2021-08-27 | 中国航空油料集团有限公司 | Data set processing method for accelerating training of target detection algorithm |
| CN113688709A (en) * | 2021-08-17 | 2021-11-23 | 长江大学 | Intelligent detection method, system, terminal and medium for wearing safety helmet |
| CN114049488A (en) * | 2022-01-07 | 2022-02-15 | 济南和普威视光电技术有限公司 | Multi-dimensional information fusion remote weak and small target detection method and terminal |
| CN114049627A (en) * | 2022-01-11 | 2022-02-15 | 浙江华是科技股份有限公司 | Ship board detection method, device and system capable of intelligently selecting scaling dimension |
| CN111191621B (en) * | 2020-01-03 | 2024-06-28 | 北京同方软件有限公司 | Rapid and accurate identification method for multi-scale target in large-focal-length monitoring scene |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107169421A (en) * | 2017-04-20 | 2017-09-15 | 华南理工大学 | A kind of car steering scene objects detection method based on depth convolutional neural networks |
| CN107609465A (en) * | 2017-07-25 | 2018-01-19 | 北京联合大学 | A kind of multi-dimension testing method for Face datection |
-
2018
- 2018-01-23 CN CN201810065807.1A patent/CN108460403A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107169421A (en) * | 2017-04-20 | 2017-09-15 | 华南理工大学 | A kind of car steering scene objects detection method based on depth convolutional neural networks |
| CN107609465A (en) * | 2017-07-25 | 2018-01-19 | 北京联合大学 | A kind of multi-dimension testing method for Face datection |
Non-Patent Citations (4)
| Title |
|---|
| JOSEPH REDMON 等: "YOLO9000: Better, Faster, Stronger", 《2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR)》 * |
| NAVANEETH BODLA 等: "Soft-NMS – Improving Object Detection With One Line of Code", 《2017 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV)》 * |
| PEIYUN HU 等: "Finding Tiny Faces", 《2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR)》 * |
| TSUNG-YI LIN 等: "Feature Pyramid Networks for Object Detection", 《2017 IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION (CVPR)》 * |
Cited By (77)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109363614A (en) * | 2018-08-29 | 2019-02-22 | 合肥德易电子有限公司 | Intelligent integral robot cavity mirror system with high definition video enhancing processing function |
| CN109363614B (en) * | 2018-08-29 | 2020-09-01 | 合肥德易电子有限公司 | Intelligent integrated robot cavity mirror system with high-definition video enhancement processing function |
| CN109241902A (en) * | 2018-08-30 | 2019-01-18 | 北京航空航天大学 | A kind of landslide detection method based on multi-scale feature fusion |
| CN109308458B (en) * | 2018-08-31 | 2022-03-15 | 电子科技大学 | Method for improving small target detection precision based on characteristic spectrum scale transformation |
| CN109308458A (en) * | 2018-08-31 | 2019-02-05 | 电子科技大学 | A method of small target deteection precision is promoted based on characteristic spectrum change of scale |
| CN109343701A (en) * | 2018-09-03 | 2019-02-15 | 电子科技大学 | A kind of intelligent human-machine interaction method based on dynamic hand gesture recognition |
| CN109255320B (en) * | 2018-09-03 | 2020-09-25 | 电子科技大学 | Improved non-maximum suppression method |
| CN109255320A (en) * | 2018-09-03 | 2019-01-22 | 电子科技大学 | A kind of improved non-maxima suppression method |
| CN109255352A (en) * | 2018-09-07 | 2019-01-22 | 北京旷视科技有限公司 | Object detection method, apparatus and system |
| CN109255352B (en) * | 2018-09-07 | 2021-06-22 | 北京旷视科技有限公司 | Target detection method, device and system |
| CN109472193A (en) * | 2018-09-21 | 2019-03-15 | 北京飞搜科技有限公司 | Method for detecting human face and device |
| CN109522958A (en) * | 2018-11-16 | 2019-03-26 | 中山大学 | Based on the depth convolutional neural networks object detection method merged across scale feature |
| CN109711427A (en) * | 2018-11-19 | 2019-05-03 | 深圳市华尊科技股份有限公司 | Object detection method and Related product |
| CN109598290A (en) * | 2018-11-22 | 2019-04-09 | 上海交通大学 | A kind of image small target detecting method combined based on hierarchical detection |
| CN111241893A (en) * | 2018-11-29 | 2020-06-05 | 阿里巴巴集团控股有限公司 | Identification recognition method, device and system |
| CN111241893B (en) * | 2018-11-29 | 2023-06-16 | 阿里巴巴集团控股有限公司 | Identification recognition method, device and system |
| CN109658412A (en) * | 2018-11-30 | 2019-04-19 | 湖南视比特机器人有限公司 | It is a kind of towards de-stacking sorting packing case quickly identify dividing method |
| CN109635717A (en) * | 2018-12-10 | 2019-04-16 | 天津工业大学 | A kind of mining pedestrian detection method based on deep learning |
| CN109615016B (en) * | 2018-12-20 | 2021-06-22 | 北京理工大学 | Target detection method of convolutional neural network based on pyramid input gain |
| CN109615016A (en) * | 2018-12-20 | 2019-04-12 | 北京理工大学 | A kind of object detection method of the convolutional neural networks based on pyramid input gain |
| CN111368600A (en) * | 2018-12-26 | 2020-07-03 | 北京眼神智能科技有限公司 | Method and device for detecting and identifying remote sensing image target, readable storage medium and equipment |
| CN111368600B (en) * | 2018-12-26 | 2023-10-31 | 北京眼神智能科技有限公司 | Remote sensing image target detection and identification method and device, readable storage medium and equipment |
| CN109815868B (en) * | 2019-01-15 | 2022-02-01 | 腾讯科技（深圳）有限公司 | Image target detection method and device and storage medium |
| CN109815868A (en) * | 2019-01-15 | 2019-05-28 | 腾讯科技（深圳）有限公司 | A kind of image object detection method, device and storage medium |
| CN109886147A (en) * | 2019-01-29 | 2019-06-14 | 电子科技大学 | A kind of more attribute detection methods of vehicle based on the study of single network multiple-task |
| CN109816671B (en) * | 2019-01-31 | 2021-09-24 | 深兰科技（上海）有限公司 | Target detection method, device and storage medium |
| CN109816036B (en) * | 2019-01-31 | 2021-08-27 | 北京字节跳动网络技术有限公司 | Image processing method and device |
| CN109816036A (en) * | 2019-01-31 | 2019-05-28 | 北京字节跳动网络技术有限公司 | Image processing method and device |
| CN109816671A (en) * | 2019-01-31 | 2019-05-28 | 深兰科技（上海）有限公司 | A kind of object detection method, device and storage medium |
| CN110046550A (en) * | 2019-03-14 | 2019-07-23 | 中山大学 | Pedestrian's Attribute Recognition system and method based on multilayer feature study |
| CN110046550B (en) * | 2019-03-14 | 2021-07-13 | 中山大学 | Pedestrian attribute identification system and method based on multilayer feature learning |
| CN109978035A (en) * | 2019-03-18 | 2019-07-05 | 西安电子科技大学 | Pedestrian detection method based on improved k-means and loss function |
| CN109978035B (en) * | 2019-03-18 | 2021-04-02 | 西安电子科技大学 | Pedestrian detection method based on improved k-means and loss function |
| CN111738036B (en) * | 2019-03-25 | 2023-09-29 | 北京四维图新科技股份有限公司 | Image processing method, device, equipment and storage medium |
| CN111738036A (en) * | 2019-03-25 | 2020-10-02 | 北京四维图新科技股份有限公司 | Image processing method, device, equipment and storage medium |
| CN110110599A (en) * | 2019-04-03 | 2019-08-09 | 天津大学 | A kind of Remote Sensing Target detection method based on multi-scale feature fusion |
| CN110110599B (en) * | 2019-04-03 | 2023-05-09 | 天津大学 | Remote sensing image target detection method based on multi-scale feature fusion |
| CN110189307A (en) * | 2019-05-14 | 2019-08-30 | 慧影医疗科技（北京）有限公司 | A kind of pulmonary nodule detection method and system based on multi-model fusion |
| CN110321864A (en) * | 2019-07-09 | 2019-10-11 | 西北工业大学 | Remote sensing images explanatory note generation method based on multiple dimensioned cutting mechanism |
| CN110503098A (en) * | 2019-08-29 | 2019-11-26 | 西安电子科技大学 | A kind of object detection method and equipment of quick real-time lightweight |
| CN110738208A (en) * | 2019-10-08 | 2020-01-31 | 创新奇智(重庆)科技有限公司 | efficient scale-normalized target detection training method |
| CN110796039A (en) * | 2019-10-15 | 2020-02-14 | 北京达佳互联信息技术有限公司 | Face flaw detection method and device, electronic equipment and storage medium |
| CN110796360A (en) * | 2019-10-24 | 2020-02-14 | 吉林化工学院 | Fixed traffic detection source multi-scale data fusion method |
| CN110909615A (en) * | 2019-10-28 | 2020-03-24 | 西安交通大学 | Target detection method based on multi-scale input mixed perception neural network |
| CN111767934B (en) * | 2019-10-31 | 2023-11-03 | 杭州海康威视数字技术股份有限公司 | Image recognition method and device and electronic equipment |
| CN111767934A (en) * | 2019-10-31 | 2020-10-13 | 杭州海康威视数字技术股份有限公司 | Image identification method and device and electronic equipment |
| CN110991359A (en) * | 2019-12-06 | 2020-04-10 | 重庆市地理信息和遥感应用中心(重庆市测绘产品质量检验测试中心) | Satellite image target detection method based on multi-scale depth convolution neural network |
| CN111144248A (en) * | 2019-12-16 | 2020-05-12 | 上海交通大学 | People counting method, system and medium based on ST-FHCD network model |
| CN111144248B (en) * | 2019-12-16 | 2024-02-27 | 上海交通大学 | People counting method, system and medium based on ST-FHCD network model |
| CN111582012A (en) * | 2019-12-24 | 2020-08-25 | 珠海大横琴科技发展有限公司 | Method and device for detecting small target ship |
| CN111127458A (en) * | 2019-12-27 | 2020-05-08 | 深圳力维智联技术有限公司 | Target detection method and device based on image pyramid and storage medium |
| CN111127458B (en) * | 2019-12-27 | 2023-06-09 | 深圳力维智联技术有限公司 | Image pyramid-based target detection method, device and storage medium |
| CN111191621B (en) * | 2020-01-03 | 2024-06-28 | 北京同方软件有限公司 | Rapid and accurate identification method for multi-scale target in large-focal-length monitoring scene |
| CN111191621A (en) * | 2020-01-03 | 2020-05-22 | 北京同方软件有限公司 | Rapid and accurate identification method for multi-scale target under large-focus monitoring scene |
| CN111047655B (en) * | 2020-01-10 | 2024-05-14 | 北京盛开互动科技有限公司 | High-definition camera cloth defect detection method based on convolutional neural network |
| CN111047655A (en) * | 2020-01-10 | 2020-04-21 | 北京盛开互动科技有限公司 | High-definition camera cloth defect detection method based on convolutional neural network |
| CN111260615A (en) * | 2020-01-13 | 2020-06-09 | 重庆交通大学 | Laser and machine vision fusion-based method for detecting apparent diseases of unmanned aerial vehicle bridge |
| CN111986255B (en) * | 2020-09-07 | 2024-04-09 | 凌云光技术股份有限公司 | Multi-scale anchor initializing method and device of image detection model |
| CN111986255A (en) * | 2020-09-07 | 2020-11-24 | 北京凌云光技术集团有限责任公司 | Multi-scale anchor initialization method and device of image detection model |
| CN112418208B (en) * | 2020-12-11 | 2022-09-16 | 华中科技大学 | Tiny-YOLO v 3-based weld film character recognition method |
| CN112418208A (en) * | 2020-12-11 | 2021-02-26 | 华中科技大学 | Tiny-YOLO v 3-based weld film character recognition method |
| CN112580721A (en) * | 2020-12-19 | 2021-03-30 | 北京联合大学 | Target key point detection method based on multi-resolution feature fusion |
| CN112580721B (en) * | 2020-12-19 | 2023-10-24 | 北京联合大学 | Target key point detection method based on multi-resolution feature fusion |
| CN112560732A (en) * | 2020-12-22 | 2021-03-26 | 电子科技大学中山学院 | Multi-scale feature extraction network and feature extraction method thereof |
| CN112560732B (en) * | 2020-12-22 | 2023-07-04 | 电子科技大学中山学院 | Feature extraction method of multi-scale feature extraction network |
| CN112528976A (en) * | 2021-02-09 | 2021-03-19 | 北京世纪好未来教育科技有限公司 | Text detection model generation method and text detection method |
| CN113222916A (en) * | 2021-04-28 | 2021-08-06 | 北京百度网讯科技有限公司 | Method, apparatus, device and medium for detecting image using target detection model |
| CN113222916B (en) * | 2021-04-28 | 2023-08-18 | 北京百度网讯科技有限公司 | Method, apparatus, device and medium for detecting image using object detection model |
| CN112906677A (en) * | 2021-05-06 | 2021-06-04 | 南京信息工程大学 | Pedestrian target detection and re-identification method based on improved SSD (solid State disk) network |
| CN113221752A (en) * | 2021-05-13 | 2021-08-06 | 北京惠朗时代科技有限公司 | Multi-template matching-based multi-scale character accurate identification method |
| CN113313213B (en) * | 2021-07-28 | 2021-11-19 | 中国航空油料集团有限公司 | Data set processing method for accelerating training of target detection algorithm |
| CN113313213A (en) * | 2021-07-28 | 2021-08-27 | 中国航空油料集团有限公司 | Data set processing method for accelerating training of target detection algorithm |
| CN113688709B (en) * | 2021-08-17 | 2023-12-05 | 广东海洋大学 | Intelligent detection method, system, terminal and medium for wearing safety helmet |
| CN113688709A (en) * | 2021-08-17 | 2021-11-23 | 长江大学 | Intelligent detection method, system, terminal and medium for wearing safety helmet |
| CN114049488A (en) * | 2022-01-07 | 2022-02-15 | 济南和普威视光电技术有限公司 | Multi-dimensional information fusion remote weak and small target detection method and terminal |
| CN114049627B (en) * | 2022-01-11 | 2022-04-08 | 浙江华是科技股份有限公司 | Ship board detection method, device and system capable of intelligently selecting scaling dimension |
| CN114049627A (en) * | 2022-01-11 | 2022-02-15 | 浙江华是科技股份有限公司 | Ship board detection method, device and system capable of intelligently selecting scaling dimension |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108460403A (en) | The object detection method and system of multi-scale feature fusion in a kind of image | |
| Luo et al. | Traffic sign recognition using a multi-task convolutional neural network | |
| CN105574550B (en) | A kind of vehicle identification method and device | |
| CN103514456B (en) | Image classification method and device based on compressed sensing multi-core learning | |
| CN108304873A (en) | Object detection method based on high-resolution optical satellite remote-sensing image and its system | |
| CN110310264A (en) | A kind of large scale object detection method, device based on DCNN | |
| CN109241913A (en) | In conjunction with the ship detection method and system of conspicuousness detection and deep learning | |
| CN108416250A (en) | Demographic method and device | |
| CN109543632A (en) | A kind of deep layer network pedestrian detection method based on the guidance of shallow-layer Fusion Features | |
| CN114220035A (en) | Rapid pest detection method based on improved YOLO V4 | |
| CN108960404B (en) | Image-based crowd counting method and device | |
| CN104166841A (en) | Rapid detection identification method for specified pedestrian or vehicle in video monitoring network | |
| CN111783589B (en) | Complex scene crowd counting method based on scene classification and multi-scale feature fusion | |
| CN108764096B (en) | Pedestrian re-identification system and method | |
| CN110263712A (en) | A kind of coarse-fine pedestrian detection method based on region candidate | |
| CN106778687A (en) | Method for viewing points detecting based on local evaluation and global optimization | |
| CN110991274B (en) | Pedestrian tumbling detection method based on Gaussian mixture model and neural network | |
| CN114758362B (en) | Clothing changing pedestrian re-identification method based on semantic perception attention and visual shielding | |
| CN113762009B (en) | Crowd counting method based on multi-scale feature fusion and double-attention mechanism | |
| CN112270331A (en) | Improved billboard detection method based on YOLOV5 | |
| CN108681696A (en) | Merge the infrared target robustness recognition methods of multiple features dimensionality reduction and transfer learning | |
| CN113158943A (en) | Cross-domain infrared target detection method | |
| CN110008861A (en) | A kind of recognition methods again of the pedestrian based on global and local feature learning | |
| CN109086659A (en) | A kind of Human bodys' response method and apparatus based on multimode road Fusion Features | |
| Zang et al. | Traffic lane detection using fully convolutional neural network |
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 |
Application publication date: 20180828 |
|
| RJ01 | Rejection of invention patent application after publication |