CN106778705A - A kind of pedestrian's individuality dividing method and device - Google Patents
A kind of pedestrian's individuality dividing method and device Download PDFInfo
- Publication number
- CN106778705A CN106778705A CN201710065013.0A CN201710065013A CN106778705A CN 106778705 A CN106778705 A CN 106778705A CN 201710065013 A CN201710065013 A CN 201710065013A CN 106778705 A CN106778705 A CN 106778705A
- Authority
- CN
- China
- Prior art keywords
- coarseness
- segmentation
- training
- parted pattern
- humanoid profile
- 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.)
- Granted
Links
- 238000000034 method Methods 0.000 title claims abstract description 23
- 230000011218 segmentation Effects 0.000 claims abstract description 126
- 230000000903 blocking effect Effects 0.000 claims abstract description 52
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 30
- 238000003475 lamination Methods 0.000 claims description 19
- 238000010606 normalization Methods 0.000 claims description 18
- 238000011478 gradient descent method Methods 0.000 claims description 10
- 230000007935 neutral effect Effects 0.000 claims description 3
- 230000001537 neural effect Effects 0.000 claims 3
- 238000010586 diagram Methods 0.000 description 2
- 241001269238 Data Species 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 235000012054 meals Nutrition 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 238000005192 partition Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/52—Surveillance or monitoring of activities, e.g. for recognising suspicious objects
- G06V20/53—Recognition of crowd images, e.g. recognition of crowd congestion
-
- 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
- G06N3/047—Probabilistic or stochastic networks
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/10—Character recognition
- G06V30/19—Recognition using electronic means
- G06V30/192—Recognition using electronic means using simultaneous comparisons or correlations of the image signals with a plurality of references
- G06V30/194—References adjustable by an adaptive method, e.g. learning
Abstract
The present invention discloses a kind of pedestrian's individuality dividing method and device, and methods described includes:Pedestrian's segmentation is carried out to pending image using training in advance good coarseness humanoid profile parted pattern, blocking segmentation result is obtained;The blocking segmentation result includes multiple blocks for being labeled as background and prospect, it is described to pending image in be labeled as the block of background not comprising pedestrian's main body, and be labeled as the parts of images of the block comprising pedestrian's main body of prospect;The background image in blocking pedestrian's segmentation result corresponding part described in the pending image is removed, coarseness segmentation image is obtained;The coarseness segmentation image is input into the good fine granularity humanoid profile parted pattern of training in advance;The good individual pedestrian's segmentation result of fine granularity humanoid profile parted pattern output of the training in advance;Wherein, the coarseness humanoid profile parted pattern and fine granularity humanoid profile parted pattern are trained by full convolutional neural networks and obtained.
Description
Technical field
The present invention relates to computer vision, mode identification technology, more particularly to based on the thick of full convolutional neural networks
Pedestrian's individuality dividing method and device that fine granularity is combined.
Background technology
Pedestrian's individuality segmentation problem is one of field sixty-four dollar questions such as scene understanding, living things feature recognition.It is most of
Can not contain other pedestrians in conventional pedestrian's dividing method requirement background, segmentation knot is obtained with the difference of environment by distinguishing human body
Really.However, the situation that pedestrian is blocked mutually in actual scene is largely present, now conventional pedestrian's dividing method cannot be expired
The result of meaning.This problem can partly be solved by way of pedestrian is split and is combined individuality detection, but individuality is detected
It is time-consuming serious, even and if be accurately detected a body position in many cases, still due in region comprising multiple human body informations
Perfect individuality segmentation result cannot be obtained.The method that the thickness granularity that this method is proposed is combined can preferably solve this and ask
Topic.
The content of the invention
The purpose of the present invention is directed to prior art and is expert at the problem that runs into individual human segmentation, by coarseness segmentation mould
The mode that type is combined with fine granularity parted pattern, using other pedestrians occurred in Coarse grain model shielding background, in this base
Fine segmentation is carried out with finely granular access control on plinth, the segmentation result of individuality is obtained.First with humanoid figure's picture of a large amount of tape labels
Train the coarseness humanoid profile parted pattern of the full convolutional neural networks of multilayer;Then mould is split using the coarseness humanoid profile
Type obtains the humanoid segmentation result of all images, and (background area subduction on image will be corresponded into picture according to the segmentation result
Element sets to 0), in this, as the input of fine granularity parted pattern;Finally by the use of the image of background is shielded as input, fine is humanoid
Mark trains the humanoid parted pattern of fine granularity as supervision message.
To achieve the above object, first aspect present invention provides a kind of pedestrian individuality dividing method, including:
Pedestrian's segmentation is carried out to pending image using training in advance good coarseness humanoid profile parted pattern, area is obtained
Block segmentation result;The blocking segmentation figure picture includes multiple blocks for being labeled as background and prospect, is labeled as background
Block in do not include pedestrian's main body, and be labeled as the parts of images comprising pedestrian's main body in the block of prospect;
The background image in blocking pedestrian's segmentation result corresponding part described in the pending image is removed, obtains thick
Granulometric image;
The coarseness segmentation image is input into the good fine granularity humanoid profile parted pattern of training in advance;
Wherein, the coarseness humanoid profile parted pattern and fine granularity humanoid profile parted pattern are by full convolution god
Obtained through network training.
Wherein, the corresponding first full convolutional neural networks of the coarseness humanoid profile parted pattern include multilayer convolutional layer
With one layer of warp lamination;The corresponding second full convolutional neural networks of the fine granularity humanoid profile parted pattern include multilayer convolution
Layer and multilayer warp lamination, the structure that the multilayer convolutional layer is centrosymmetric with the multilayer warp lamination are combined into funnel
Shape.
The methods described also training step including the coarseness humanoid profile parted pattern, including:
The training sample with pedestrian's mark that training data is concentrated is carried out into blocking treatment, the area of training sample is obtained
Block result;
Unified size will be normalized to for the training sample trained, it is then that the training sample feeding after normalization is thick
The corresponding first full convolutional neural networks of granularity humanoid profile parted pattern;
The block of block segmentation result and corresponding training sample that the coarseness humanoid profile parted pattern is exported
Change result is compared and obtains predicated error;
Back-propagation algorithm and stochastic gradient descent method is used to reduce predicated error to train the coarseness humanoid profile
The corresponding first full convolutional neural networks of parted pattern, obtain final coarseness humanoid profile and split by successive ignition training
Model.
The methods described also training step including the fine granularity humanoid profile parted pattern, including:
The training sample with pedestrian's mark that training data is concentrated is input into the coarseness humanoid profile for training point
Model is cut, blocking segmentation result is obtained;
By the background image subduction in blocking segmentation result corresponding part described in the training sample, coarseness is obtained
Segmentation figure picture;
By coarseness segmentation image normalization to unified size;
By the corresponding second full convolution of coarseness segmentation image feeding fine granularity humanoid profile parted pattern after normalization
In neutral net;
By the fine segmentation of the fine segmentation result of the described second full convolutional neural networks output and corresponding training sample
Mark result is compared and obtains the second predicated error;
Back-propagation algorithm and stochastic gradient descent method is used to reduce the second predicated error to train the fine granularity humanoid
The corresponding second full convolutional neural networks of contours segmentation model, final fine granularity humanoid profile is obtained by successive ignition training
Parted pattern.
Wherein, the supervision message of the coarseness humanoid profile parted pattern is the dividing mark for carrying out blocking treatment,
For the background in masked images.
Second aspect present invention provides a kind of pedestrian individuality segmenting device, including:
Blocking segmentation module, is configured to, with the good coarseness humanoid profile parted pattern of training in advance to pending
Image carries out pedestrian's segmentation, obtains blocking segmentation figure picture;The blocking segmentation figure picture include multiple be labeled as background and
The block of prospect, is labeled as in the block of background not comprising pedestrian's main body, and is labeled as in the block of prospect comprising pedestrian
The parts of images of main body;
Background removal module, is configured as blocking pedestrian's segmentation result correspondence portion described in the removal pending image
Background image in point, obtains coarseness segmentation image;
Fine segmentation module, is configured as just described coarseness segmentation image and is input into the good fine granularity people of training in advance
Shape contours segmentation model;
Wherein, the coarseness humanoid profile parted pattern and fine granularity humanoid profile parted pattern are by full convolution god
Obtained through network training.
Wherein, the corresponding first full convolutional neural networks of the coarseness humanoid profile parted pattern include multilayer convolutional layer
With one layer of warp lamination;The corresponding second full convolutional neural networks of the fine granularity humanoid profile parted pattern include multilayer convolution
Layer and multilayer warp lamination, the structure that the multilayer convolutional layer is centrosymmetric with the multilayer warp lamination are combined into funnel
Shape.
The described device also training module including the coarseness humanoid profile parted pattern, including:
Mark submodule, the training sample with pedestrian's mark for being configured as concentrating training data carries out blocking place
Reason, obtains the blocking result of training sample;
First normalization submodule, is configured as that unified size will be normalized to for the training sample trained;
First training submodule, is configured as the training sample after normalization sending into coarseness humanoid profile parted pattern
Corresponding first full convolutional neural networks;
First comparison sub-module, is configured as the block segmentation result for exporting the coarseness humanoid profile parted pattern
It is compared with the blocking result of corresponding training sample and obtains predicated error;
First iteration submodule, is configured as reducing predicated error using back-propagation algorithm and stochastic gradient descent method
To train the corresponding first full convolutional neural networks of the coarseness humanoid profile parted pattern, obtained most by successive ignition training
Whole coarseness humanoid profile parted pattern.
Described device, also including the training module of the fine granularity humanoid profile parted pattern, including:
Blocking submodule, the training sample with pedestrian's mark for being configured as concentrating training data is input into training
Good coarseness humanoid profile parted pattern, obtains blocking segmentation result;
Background removal submodule, is configured as in blocking segmentation result corresponding part described in the training sample
Background image is reduced, and obtains coarseness segmentation image;
Second normalization submodule, is configured as coarseness segmentation image normalization to unified size;
Second training submodule, is configured as the coarseness segmentation image feeding fine granularity humanoid profile point after normalization
In cutting the corresponding second full convolutional neural networks of model;
Second comparison sub-module, be configured as by the fine segmentation result of the described second full convolutional neural networks output with it is right
The fine segmentation mark result of the training sample answered is compared and obtains the second predicated error;
Secondary iteration submodule, is configured as reducing the second prediction using back-propagation algorithm and stochastic gradient descent method
Error is trained with training the corresponding second full convolutional neural networks of the fine granularity humanoid profile parted pattern by successive ignition
To final fine granularity humanoid profile parted pattern.
Wherein, the supervision message of the coarseness humanoid profile parted pattern is the dividing mark for carrying out blocking treatment,
For the background in masked images.
Pedestrian's individuality dividing method that the thickness granularity based on full convolutional neural networks that the present invention is provided is combined is using deep
Learning art is spent to be respectively trained coarseness segmentation model and finely granular access control, and using the result shielding of coarseness segmentation model
Part background, can improve the accuracy rate of humanoid segmentation, be particularly suited for the situation containing other pedestrians in background.It is of the invention
Above-mentioned technical proposal need to be used to remove in combination with the result of thickness granulometric model, the wherein result of coarseness segmentation model
Background block in image, and in this, as the input of fine granularity parted pattern, the difficulty of fine granularity segmentation can be substantially reduced, change
Kind segmentation effect;Thickness granulometric model is full convolutional neural networks, and only includes convolutional layer and full convolution Rotating fields,
With simple structure, the few advantage of parameter, therefore the speed of service is very fast;Wherein coarseness segmentation network only includes one layer of deconvolution
Layer, be able to can be predicted fine with the segmentation result of prediction block, fine granularity segmentation network funnel-shaped structure symmetrical before and after being
Segmentation result;The supervision message of the coarseness segmentation model for being proposed is the dividing mark for carrying out blocking treatment, by instruction
White silk can effectively in masked images background, especially other pedestrians in background finally only retain containing pedestrian
Region
Brief description of the drawings
Fig. 1 is training data and labeling method schematic diagram in the present invention;
Fig. 2 is pedestrian's individuality dividing method schematic flow sheet in one embodiment of the invention;
Fig. 3 is coarseness segmentation model structure schematic diagram in one embodiment of the invention;
Fig. 4 is fine granularity parted pattern structural representation in one embodiment of the invention.
Specific embodiment
Below by drawings and Examples, technical scheme is described in further detail.
One embodiment of the invention proposes a kind of pedestrian's individuality dividing method.The method includes:
Pedestrian's segmentation is carried out to pending image using training in advance good coarseness humanoid profile parted pattern, area is obtained
Block segmentation result;The blocking segmentation figure picture includes multiple blocks for being labeled as background and prospect, is labeled as background
Block in do not include pedestrian's main body, and be labeled as the parts of images comprising pedestrian's main body in the block of prospect;
The background image in blocking pedestrian's segmentation result corresponding part described in the pending image is removed, obtains thick
Granulometric image;
The coarseness segmentation image is input into the good fine granularity humanoid profile parted pattern of training in advance;
The good individual pedestrian's segmentation result of fine granularity humanoid profile parted pattern output of the training in advance.
In one embodiment, the blocking segmentation result is that pending image is divided into multiple size identical areas
Block, and each block is labeled as background block or prospect block, pedestrian is not included in the corresponding image of the background block
Subject image, and the parts of images comprising pedestrian's main body in the corresponding image of the prospect block, as shown in figure 1, d is block point
Cut result, e is blocking segmentation figure picture, i.e., the corresponding blocking segmentation figure picture of blocking segmentation result.
In one embodiment, the coarseness humanoid profile parted pattern and the fine granularity humanoid profile parted pattern are equal
It is full convolutional neural networks, i.e., only includes convolutional layer and full convolution Rotating fields, with simple structure, the few advantage of parameter, because
This speed of service is very fast.
In one embodiment, the coarseness humanoid profile parted pattern includes multilayer convolutional layer and one layer of warp lamination,
For the segmentation result of prediction block, and fine granularity humanoid profile parted pattern includes multilayer convolutional layer and multilayer deconvolution
Symmetrical structure centered on layer, its conventional part and deconvolution part, is combined into funnel shaped, i.e., described fine granularity humanoid profile
The most intermediate layer of parted pattern is convolutional layer, and first half is convolutional layer, and latter half is meal convolutional layer, first half and
The structure that latter half is centrosymmetric.
In one embodiment, the supervision message of the coarseness humanoid profile parted pattern is to carry out dividing for blocking treatment
Mark is cut, background that can effectively in masked images by training, especially other pedestrians in background, finally only reservation contains
There is the region of pedestrian.
The present invention Coarse grain model segmentation on the basis of, finely granular access control by the symmetrical full convolutional network of infundibulate,
Very fine pedestrian's individuality segmentation result can be obtained.The method has extremely strong Shandong for background change various in image
Rod, can preferably solve the humanoid segmentation problem under the mutual circumstance of occlusion of many pedestrians
Below by taking certain large-scale humanoid partition data storehouse as an example, altogether comprising 5000 pedestrian images and corresponding humanoid segmentation mark
Note.
Fig. 2 is the flow chart of pedestrian's individuality dividing method of the present invention, as illustrated, the present invention specifically includes following steps:
5000 pedestrian's flag datas in data set are carried out blocking treatment, as shown in Figure 1, first by step S0
Pedestrian's dividing mark image uniform is divided into 10 × 5 blocks, then according to whether true containing pedestrian's dividing mark in block
The fixed block is prospect or background block, has thus obtained 10 × 5 blocking dividing mark, totally 5000 pairs of pedestrian images
With block dividing mark;
Step S1, will be normalized to unified size (50 × 25 pixel), then by the figure for the pedestrian image trained
As one full convolutional neural networks (i.e. coarseness segmentation network) of feeding, the network contains several layers of convolutional layer and warp lamination, tool
Body structure as shown in figure 3, include 4 convolutional layers and 1 warp lamination altogether.First convolutional layer contains 48 wave filter (sizes
For 3 × 3), step-length is 2;Likewise, second and third, containing 96/96/128 wave filter, (size is 3 respectively for four convolutional layers
× 3), step-length is 2, and layer 5 is warp lamination, and containing 1 wave filter (size is 10 × 5), step-length is 1, and its output is
Coarseness segmentation result.
Step S2, expresses in last layer of output image of the coarseness segmentation network, i.e., (size is block segmentation result
10×5);
Step S3, the block segmentation result of above-mentioned output is entered with corresponding blocking dividing mark (as shown in d in Fig. 1)
Go to compare and obtain predicated error, can now compare the predicated error of each point in 10 × 5 regions and sue for peace as final prediction
Error;
Step S4, uses back-propagation algorithm and stochastic gradient descent method to reduce predicated error to train the coarseness point
Network is cut, preferable coarseness segmentation model is obtained by successive ignition training, take around the iteration of 10,000 time or so, can
Lost with further reducing error by adjusting the learning rate of weight, deconditioning when error no longer reduces, now slightly
Granulometric model training is completed;
Step S5, the Coarse grain model that pedestrian image (size is 50 × 25) the input training after normalization is completed, can
Blocking segmentation result (size is 10 × 5) is obtained, according to the segmentation result by not normalized original pedestrian image
Background area subduction (will respective pixel set to 0), obtains removing the pedestrian image of background, as shown in Figure 1;
Step S6, the pedestrian image of the removal background that S5 is obtained is normalized to unified size (such as 150 × 75 pixels),
Then by the image send into a full convolutional neural networks (i.e. fine granularity segmentation network), the network include several layers of convolutional layer with it is anti-
Convolutional layer, its conventional part is usually symmetrical structure with deconvolution part, is combined into funnel shaped, as shown in figure 4, including 4 altogether
Individual convolutional layer and 3 warp laminations.First convolutional layer contains 48 wave filters (size is 3 × 3), and step-length is 2;Likewise,
Second and third, four convolutional layers contain 64/96/128 wave filter (size is 3 × 3) respectively, step-length is 2;Layer 5 is
Warp lamination, symmetrical with third layer convolutional layer, containing 96 wave filters (size is 3 × 3), step-length is 2;Layer 6 is deconvolution
Layer, symmetrical with second layer convolutional layer, containing 64 wave filters (size is 3 × 3), step-length is 2;Layer 7 is warp lamination, with
Ground floor convolutional layer is symmetrical, and containing 1 wave filter (size is 3 × 3), step-length is 2, its subdivision for being output as 150 × 75 sizes
Cut result;
Step S7, in last layer of output image expression of fine granularity segmentation network, i.e., (size is fine segmentation result
150×75);
Step S8, by the fine segmentation result of above-mentioned output with it is corresponding normalization dividing mark (size be 150 × 75,
As shown in b in Fig. 1) it is compared and obtains predicated error, the error is 150 × 75 summations of the error of pixel;It is described to return
One change dividing mark is after the original sample with accurate dividing mark is normalized into fine segmentation result sizes (50 × 75)
Result;
Step S9, uses back-propagation algorithm and stochastic gradient descent method to reduce predicated error to train the fine granularity point
Network is cut, final fine granularity parted pattern is obtained by successive ignition training;Because network size is larger, it usually needs 100,
The iteration of 000 time or so, can further reduce error loss, until error no longer reduces by adjusting the learning rate of weight
When deconditioning, now fine granularity parted pattern training complete;
Step S10, is tested using the thickness granulometric model for training.First by the figure containing pedestrian to be tested
As being normalized to 50 × 25 pixels, send into coarseness segmentation model, obtain coarseness segmentation result (i.e. block segmentation result, greatly
It is small be 10 × 5);
Step S11, is reduced the background area in original pedestrian image using the block segmentation result for obtaining in S10
Then (will respective pixel set to 0) be normalized to unified size (150 × 75 to obtain removing the pedestrian image of background
Pixel), finally by image feeding fine granularity segmentation network;
Step S12, now, the pedestrian's segmentation result that can be become more meticulous from the output end of fine granularity segmentation network.
Particular embodiments described above, has been carried out further in detail to the purpose of the present invention, technical scheme and beneficial effect
Describe in detail bright, it should be understood that the foregoing is only specific embodiment of the invention, be not intended to limit the invention, it is all
Within the spirit and principles in the present invention, any modification, equivalent substitution and improvements done etc. should be included in protection of the invention
Within the scope of.
Claims (10)
1. a kind of pedestrian individuality dividing method, including:
Pedestrian's segmentation is carried out to pending image using training in advance good coarseness humanoid profile parted pattern, obtains blocking
Segmentation result;The blocking segmentation result includes multiple blocks for being labeled as background and prospect, described to pending figure
The block of background is labeled as in not comprising pedestrian's main body, and is labeled as the part figure of the block comprising pedestrian's main body of prospect
Picture;
The background image in blocking pedestrian's segmentation result corresponding part described in the pending image is removed, coarseness is obtained
Segmentation figure picture;
The coarseness segmentation image is input into the good fine granularity humanoid profile parted pattern of training in advance;
The good individual pedestrian's segmentation result of fine granularity humanoid profile parted pattern output of the training in advance;
Wherein, the coarseness humanoid profile parted pattern and fine granularity humanoid profile parted pattern pass through full convolutional Neural net
Network training is obtained.
2. the method for claim 1, wherein the corresponding first full convolution of the coarseness humanoid profile parted pattern is refreshing
Include multilayer convolutional layer and one layer of warp lamination through network;The corresponding second full convolution of the fine granularity humanoid profile parted pattern
Neutral net includes multilayer convolutional layer and multilayer warp lamination, and the multilayer convolutional layer is in center pair with the multilayer warp lamination
The structure of title, is combined into funnel shaped.
3. the method for claim 1, also including the training step of the coarseness humanoid profile parted pattern, including:
The training sample with pedestrian's mark that training data is concentrated is carried out into blocking treatment, the blocking of training sample is obtained
Result;
Unified size will be normalized to for the training sample trained, the training sample after normalization is then sent into coarseness
The corresponding first full convolutional neural networks of humanoid profile parted pattern;
The blocking place of block segmentation result and corresponding training sample that the coarseness humanoid profile parted pattern is exported
Reason result is compared and obtains predicated error;
Back-propagation algorithm and stochastic gradient descent method is used to reduce predicated error to train the coarseness humanoid profile to split
The corresponding first full convolutional neural networks of model, obtain final coarseness humanoid profile and split mould by successive ignition training
Type.
4. the method for claim 1, also including the training step of the fine granularity humanoid profile parted pattern, including:
The training sample with pedestrian's mark that training data is concentrated is input into the coarseness humanoid profile for training and splits mould
Type, obtains blocking segmentation result;
By the background image subduction in blocking segmentation result corresponding part described in the training sample, coarseness segmentation is obtained
Image;
By coarseness segmentation image normalization to unified size;
By the corresponding second full convolutional Neural of coarseness segmentation image feeding fine granularity humanoid profile parted pattern after normalization
In network;
The fine segmentation result of the described second full convolutional neural networks output is marked with the fine segmentation of corresponding training sample
Result is compared and obtains the second predicated error;
Back-propagation algorithm and stochastic gradient descent method is used to reduce the second predicated error to train the fine granularity humanoid profile
The corresponding second full convolutional neural networks of parted pattern, obtain final fine granularity humanoid profile and split by successive ignition training
Model.
5. method according to claim 1, wherein, the supervision message of the coarseness humanoid profile parted pattern is to carry out
The dividing mark of blocking treatment, for the background in masked images.
6. a kind of pedestrian individuality segmenting device, including:
Blocking segmentation module, is configured to, with the good coarseness humanoid profile parted pattern of training in advance to pending image
Pedestrian's segmentation is carried out, blocking segmentation figure picture is obtained;The blocking segmentation figure picture includes that multiple is labeled as background and prospect
Block, be labeled as in the block of background not comprising pedestrian's main body, and be labeled as in the block of prospect comprising pedestrian's main body
Parts of images;
Background removal module, is configured as described in the removal pending image in blocking pedestrian's segmentation result corresponding part
Background image, obtain coarseness segmentation image;
Fine segmentation module, is configured as just described coarseness segmentation image and is input into the training in advance humanoid wheel of good fine granularity
Wide parted pattern;
As a result output module, is configured as the good individual pedestrian point of fine granularity humanoid profile parted pattern output of the training in advance
Cut result;
Wherein, the coarseness humanoid profile parted pattern and fine granularity humanoid profile parted pattern pass through full convolutional Neural net
Network training is obtained.
7. device as claimed in claim 6, wherein, the corresponding first full convolution god of the coarseness humanoid profile parted pattern
Include multilayer convolutional layer and one layer of warp lamination through network;The corresponding second full convolution of the fine granularity humanoid profile parted pattern
Neutral net includes multilayer convolutional layer and multilayer warp lamination, and the multilayer convolutional layer is in center pair with the multilayer warp lamination
The structure of title, is combined into funnel shaped.
8. device as claimed in claim 1, also including the training module of the coarseness humanoid profile parted pattern, including:
Mark submodule, the training sample with pedestrian's mark for being configured as concentrating training data carries out blocking treatment,
Obtain the blocking result of training sample;
First normalization submodule, is configured as that unified size will be normalized to for the training sample trained;
First training submodule, is configured as the training sample feeding coarseness humanoid profile parted pattern correspondence after normalization
The first full convolutional neural networks;
First comparison sub-module, be configured as by the coarseness humanoid profile parted pattern export block segmentation result with it is right
The blocking result of the training sample answered is compared and obtains predicated error;
First iteration submodule, is configured as using back-propagation algorithm and stochastic gradient descent method to reduce predicated error to instruct
Practice the corresponding first full convolutional neural networks of the coarseness humanoid profile parted pattern, obtain final by successive ignition training
Coarseness humanoid profile parted pattern.
9. device as claimed in claim 6, also including the training module of the fine granularity humanoid profile parted pattern, including:
Blocking submodule, the training sample with pedestrian's mark for being configured as concentrating training data is input into the coarse grain for training
Degree humanoid profile parted pattern, obtains blocking segmentation result;
Background removal submodule, is configured as the background in blocking segmentation result corresponding part described in the training sample
Image subtraction, obtains coarseness segmentation image;
Second normalization submodule, is configured as coarseness segmentation image normalization to unified size;
Second training submodule, is configured as the coarseness segmentation image feeding fine granularity humanoid profile segmentation mould after normalization
In the corresponding second full convolutional neural networks of type;
Second comparison sub-module, be configured as by the fine segmentation result of the described second full convolutional neural networks output with it is corresponding
The fine segmentation mark result of training sample is compared and obtains the second predicated error;
Secondary iteration submodule, is configured as reducing the second predicated error using back-propagation algorithm and stochastic gradient descent method
To train the corresponding second full convolutional neural networks of the fine granularity humanoid profile parted pattern, obtained most by successive ignition training
Whole fine granularity humanoid profile parted pattern.
10. device according to claim 1, wherein, the supervision message of the coarseness humanoid profile parted pattern be into
The dividing mark of the blocking treatment of row, for the background in masked images.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710065013.0A CN106778705B (en) | 2017-02-04 | 2017-02-04 | Pedestrian individual segmentation method and device |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710065013.0A CN106778705B (en) | 2017-02-04 | 2017-02-04 | Pedestrian individual segmentation method and device |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106778705A true CN106778705A (en) | 2017-05-31 |
CN106778705B CN106778705B (en) | 2020-03-17 |
Family
ID=58955591
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710065013.0A Active CN106778705B (en) | 2017-02-04 | 2017-02-04 | Pedestrian individual segmentation method and device |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106778705B (en) |
Cited By (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107766890A (en) * | 2017-10-31 | 2018-03-06 | 天津大学 | The improved method that identification segment learns in a kind of fine granularity identification |
CN107944399A (en) * | 2017-11-28 | 2018-04-20 | 广州大学 | A kind of pedestrian's recognition methods again based on convolutional neural networks target's center model |
CN108198192A (en) * | 2018-01-15 | 2018-06-22 | 任俊芬 | A kind of quick human body segmentation's method of high-precision based on deep learning |
CN108510000A (en) * | 2018-03-30 | 2018-09-07 | 北京工商大学 | The detection and recognition methods of pedestrian's fine granularity attribute under complex scene |
CN108711150A (en) * | 2018-05-22 | 2018-10-26 | 电子科技大学 | A kind of end-to-end pavement crack detection recognition method based on PCA |
CN108960190A (en) * | 2018-07-23 | 2018-12-07 | 西安电子科技大学 | SAR video object detection method based on FCN Image Sequence Model |
CN109635812A (en) * | 2018-11-29 | 2019-04-16 | 中国科学院空间应用工程与技术中心 | The example dividing method and device of image |
CN109636806A (en) * | 2018-11-22 | 2019-04-16 | 浙江大学山东工业技术研究院 | A kind of three-dimensional NMR pancreas image partition method based on multistep study |
CN109993187A (en) * | 2017-12-29 | 2019-07-09 | 深圳市优必选科技有限公司 | A kind of modeling method, robot and the storage device of object category for identification |
CN110084156A (en) * | 2019-04-12 | 2019-08-02 | 中南大学 | A kind of gait feature abstracting method and pedestrian's personal identification method based on gait feature |
CN110516583A (en) * | 2019-08-21 | 2019-11-29 | 中科视语(北京)科技有限公司 | A kind of vehicle recognition methods, system, equipment and medium again |
CN110689542A (en) * | 2018-07-04 | 2020-01-14 | 清华大学 | Portrait segmentation processing method and device based on multi-stage convolution neural network |
CN110855875A (en) * | 2018-08-20 | 2020-02-28 | 珠海格力电器股份有限公司 | Method and device for acquiring background information of image |
CN111368788A (en) * | 2020-03-17 | 2020-07-03 | 北京迈格威科技有限公司 | Training method and device of image recognition model and electronic equipment |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102609686A (en) * | 2012-01-19 | 2012-07-25 | 宁波大学 | Pedestrian detection method |
CN105760835A (en) * | 2016-02-17 | 2016-07-13 | 天津中科智能识别产业技术研究院有限公司 | Gait segmentation and gait recognition integrated method based on deep learning |
CN106022221A (en) * | 2016-05-09 | 2016-10-12 | 腾讯科技(深圳)有限公司 | Image processing method and processing system |
CN106127164A (en) * | 2016-06-29 | 2016-11-16 | 北京智芯原动科技有限公司 | The pedestrian detection method with convolutional neural networks and device is detected based on significance |
CN106355188A (en) * | 2015-07-13 | 2017-01-25 | 阿里巴巴集团控股有限公司 | Image detection method and device |
-
2017
- 2017-02-04 CN CN201710065013.0A patent/CN106778705B/en active Active
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102609686A (en) * | 2012-01-19 | 2012-07-25 | 宁波大学 | Pedestrian detection method |
CN106355188A (en) * | 2015-07-13 | 2017-01-25 | 阿里巴巴集团控股有限公司 | Image detection method and device |
CN105760835A (en) * | 2016-02-17 | 2016-07-13 | 天津中科智能识别产业技术研究院有限公司 | Gait segmentation and gait recognition integrated method based on deep learning |
CN106022221A (en) * | 2016-05-09 | 2016-10-12 | 腾讯科技(深圳)有限公司 | Image processing method and processing system |
CN106127164A (en) * | 2016-06-29 | 2016-11-16 | 北京智芯原动科技有限公司 | The pedestrian detection method with convolutional neural networks and device is detected based on significance |
Non-Patent Citations (2)
Title |
---|
E. SHELHAMER等: "Fully convolutional networks for semantic segmentation", 《IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE》 * |
卢宏涛,张秦川: "深度卷积神经网络在计算机视觉中的应用研究综述", 《数据采集与处理》 * |
Cited By (19)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107766890A (en) * | 2017-10-31 | 2018-03-06 | 天津大学 | The improved method that identification segment learns in a kind of fine granularity identification |
CN107766890B (en) * | 2017-10-31 | 2021-09-14 | 天津大学 | Improved method for discriminant graph block learning in fine-grained identification |
CN107944399A (en) * | 2017-11-28 | 2018-04-20 | 广州大学 | A kind of pedestrian's recognition methods again based on convolutional neural networks target's center model |
CN109993187A (en) * | 2017-12-29 | 2019-07-09 | 深圳市优必选科技有限公司 | A kind of modeling method, robot and the storage device of object category for identification |
CN108198192A (en) * | 2018-01-15 | 2018-06-22 | 任俊芬 | A kind of quick human body segmentation's method of high-precision based on deep learning |
CN108510000A (en) * | 2018-03-30 | 2018-09-07 | 北京工商大学 | The detection and recognition methods of pedestrian's fine granularity attribute under complex scene |
CN108510000B (en) * | 2018-03-30 | 2021-06-15 | 北京工商大学 | Method for detecting and identifying fine-grained attribute of pedestrian in complex scene |
CN108711150A (en) * | 2018-05-22 | 2018-10-26 | 电子科技大学 | A kind of end-to-end pavement crack detection recognition method based on PCA |
CN108711150B (en) * | 2018-05-22 | 2022-03-25 | 电子科技大学 | End-to-end pavement crack detection and identification method based on PCA |
CN110689542A (en) * | 2018-07-04 | 2020-01-14 | 清华大学 | Portrait segmentation processing method and device based on multi-stage convolution neural network |
CN108960190A (en) * | 2018-07-23 | 2018-12-07 | 西安电子科技大学 | SAR video object detection method based on FCN Image Sequence Model |
CN108960190B (en) * | 2018-07-23 | 2021-11-30 | 西安电子科技大学 | SAR video target detection method based on FCN image sequence model |
CN110855875A (en) * | 2018-08-20 | 2020-02-28 | 珠海格力电器股份有限公司 | Method and device for acquiring background information of image |
CN109636806A (en) * | 2018-11-22 | 2019-04-16 | 浙江大学山东工业技术研究院 | A kind of three-dimensional NMR pancreas image partition method based on multistep study |
CN109635812A (en) * | 2018-11-29 | 2019-04-16 | 中国科学院空间应用工程与技术中心 | The example dividing method and device of image |
CN110084156A (en) * | 2019-04-12 | 2019-08-02 | 中南大学 | A kind of gait feature abstracting method and pedestrian's personal identification method based on gait feature |
CN110516583A (en) * | 2019-08-21 | 2019-11-29 | 中科视语(北京)科技有限公司 | A kind of vehicle recognition methods, system, equipment and medium again |
CN111368788A (en) * | 2020-03-17 | 2020-07-03 | 北京迈格威科技有限公司 | Training method and device of image recognition model and electronic equipment |
CN111368788B (en) * | 2020-03-17 | 2023-10-27 | 北京迈格威科技有限公司 | Training method and device for image recognition model and electronic equipment |
Also Published As
Publication number | Publication date |
---|---|
CN106778705B (en) | 2020-03-17 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106778705A (en) | A kind of pedestrian's individuality dividing method and device | |
CN108596277B (en) | Vehicle identity recognition method and device and storage medium | |
CN107578418B (en) | Indoor scene contour detection method fusing color and depth information | |
Bautista et al. | Convolutional neural network for vehicle detection in low resolution traffic videos | |
CN105760835B (en) | A kind of gait segmentation and Gait Recognition integral method based on deep learning | |
EP3667286A1 (en) | Tire image recognition method and tire image recognition device | |
CN108108761A (en) | A kind of rapid transit signal lamp detection method based on depth characteristic study | |
CN107571867A (en) | Method and apparatus for controlling automatic driving vehicle | |
EP3566177A1 (en) | A method and apparatus for detecting objects of interest in images | |
CN108960261A (en) | A kind of obvious object detection method based on attention mechanism | |
CN104036243B (en) | A kind of Activity recognition method based on Optic flow information | |
CN108205649B (en) | Method and device for recognizing state of driver for calling and answering | |
CN109711268B (en) | Face image screening method and device | |
CN103985130B (en) | A kind of saliency analysis method for complex texture image | |
CN109635825A (en) | Vehicle attribute detection method, device and storage medium | |
CN107944403A (en) | Pedestrian's attribute detection method and device in a kind of image | |
CN107704797B (en) | Real-time detection method, system and equipment based on pedestrians and vehicles in security video | |
CN109544564A (en) | A kind of medical image segmentation method | |
CN111062331B (en) | Image mosaic detection method and device, electronic equipment and storage medium | |
CN104766324A (en) | White blood cell segmentation method based on super pixels and anomaly detection color blood cell image | |
CN110346699A (en) | Insulator arc-over information extracting method and device based on ultraviolet image processing technique | |
CN112287912A (en) | Deep learning-based lane line detection method and device | |
CN112307853A (en) | Detection method of aerial image, storage medium and electronic device | |
CN111753805A (en) | Method and device for detecting wearing of safety helmet | |
CN109614933A (en) | A kind of motion segmentation method based on certainty fitting |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |