CN108921822A - Image object method of counting based on convolutional neural networks - Google Patents
Image object method of counting based on convolutional neural networks Download PDFInfo
- Publication number
- CN108921822A CN108921822A CN201810564162.6A CN201810564162A CN108921822A CN 108921822 A CN108921822 A CN 108921822A CN 201810564162 A CN201810564162 A CN 201810564162A CN 108921822 A CN108921822 A CN 108921822A
- Authority
- CN
- China
- Prior art keywords
- image
- density
- training
- target
- network
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30242—Counting objects in image
Abstract
The invention discloses a kind of image object method of counting based on convolutional neural networks, the feature for arriving e-learning by robust enhancement layer, with more robustness, while reducing the computation complexity of model to target deformation;And density estimation is carried out using Pyramid technology counting module, the multi-scale information for including in convolutional neural networks layered characteristic is made full use of, significantly improves computational efficiency while realizing accurate counting.In short, realizing the target accurate metering in image the present invention is based on convolutional neural networks, it can be adapted for the object count task under complex scene, computation complexity is low, practical application value with higher.
Description
Technical field
The present invention relates to based on technical field of computer vision more particularly to a kind of image object by convolutional neural networks
Counting method.
Background technique
As the high speed development and people of computer technology, network communication technology and electronic technology are to social public security
It is required that continuous improvement, the intelligent video monitoring system based on intelligent video analysis technology is widely used.As intelligence
Important content in energy field of video monitoring, object count have a large amount of application scenarios in real life, accurately estimate
The specific number of target in the picture is the key that related system processing out.In intelligent transportation system, friendship is accurately estimated
Number of vehicles in logical scene can carry out public transport management for traffic management department and provide important evidence;To the passenger flow in market
Amount is counted, and business hours and the personnel assignment in market can be instructed;To the crowd density prison of the public places such as megastore
Control can find in time security risk and provide early warning.
The target of counting load is the quantity for allowing computer accurately to estimate object of interest in image.Mainstream at present
Object count method is mainly based upon the method and density estimation method neural network based that provincial characteristics returns.Wherein, area
The method that characteristic of field returns is by establishing the regression model of foreground region image feature and destination number come direct estimation scene
In target sum, such algorithm computation complexity is lower, but has ignored the spatial position distributed intelligence of target in the scene,
It is only capable of obtaining an one-dimensional statistic, and the extraction of feature depends on the foreground segmentation effect of image, robustness is insufficient.
The method of counting of density estimation is the density profile that target to be counted is generated by the sample of handmarking, is directly learnt from picture
Mapping relations of the vegetarian refreshments feature to target density distribution map.The target density distribution map of generation had both contained complete Density Distribution
Information can obtain the target numbers of arbitrary region by areal concentration summation, while contain the sky of target in the picture
Between distributed intelligence, be the emphasis of current research.
The density estimation method for being currently based on neural network is needed mostly using multi-channel network structure extraction Analysis On Multi-scale Features.
Such as high Sheng Hua is in China Patent Publication No. CN105528589A《Single image crowd based on multiple row convolutional neural networks counts
Algorithm》Middle to use multiple row convolutional network structure extraction scene characteristic, the convolution kernel that each sub-network is used is of different sizes, passes through group
Different size of receptive field feature is closed to handle the target scale variation issue in scene;Liu Yu etc. is in China Patent Publication No.
CN107506692《A kind of dense population counting and personnel's distribution estimation method based on deep learning》In equally use multiple row
Convolutional network structure extracts Analysis On Multi-scale Features by four column depth residual error networks;Similar, Deng Tengfei etc. is disclosed in Chinese patent
Number CN107301387A《A kind of image Dense crowd method of counting based on deep learning》In pass through two column convolutional networks point
It Xue Xi not high-level characteristic and low-level feature.But in above scheme, multiple row convolutional network model parameter amount is big, computation complexity
Height is difficult the requirement for meeting practical application to efficient process.
Summary of the invention
The object of the present invention is to provide a kind of image object method of counting based on convolutional neural networks, can be not significant
In the case where increasing network extraction feature complexity, model performance is further increased.
The purpose of the present invention is what is be achieved through the following technical solutions:
A kind of image object method of counting based on convolutional neural networks, including:
Pyramid object count network is established based on convolutional neural networks;
It is marked using artificial data, the Density Distribution true value image of interesting target is established on training image;
The training image that training data is concentrated by random cropping and flip horizontal mode and corresponding Density Distribution
True value image carries out enumeration data enhancing;
Using the enhanced training image of enumeration data and target density distribution true value image as pyramid object count net
The input of network completes pyramid object count network training by continuous iteration optimization, generates pyramid object count network mould
Type;
When new images input, the image with input picture block same size is generated by sliding window mode, is sent to golden word
In tower object count network model, the density profile predicted is averaged to obtain final to the density value of lap
Output density figure, to acquire target numbers.
As seen from the above technical solution provided by the invention, convolutional neural networks are based on, the mesh in image is realized
Accurate metering is marked, can be adapted for the object count task under complex scene, computation complexity is low, practical application with higher
Value.
Detailed description of the invention
In order to illustrate the technical solution of the embodiments of the present invention more clearly, required use in being described below to embodiment
Attached drawing be briefly described, it should be apparent that, drawings in the following description are only some embodiments of the invention, for this
For the those of ordinary skill in field, without creative efforts, it can also be obtained according to these attached drawings other
Attached drawing.
Fig. 1 is a kind of process of the image object method of counting based on convolutional neural networks provided in an embodiment of the present invention
Figure;
Fig. 2 is the schematic diagram of pyramid object count network provided in an embodiment of the present invention;
Fig. 3 is the schematic diagram of target's center's point diagram provided in an embodiment of the present invention and Density Distribution true value image;
Fig. 4 is Pyramid technology counting module schematic diagram provided in an embodiment of the present invention;
Fig. 5 is that Shanghaitech-B data set provided in an embodiment of the present invention exports result schematic diagram;
Fig. 6 is that TRANCOS data set provided in an embodiment of the present invention exports result schematic diagram.
Specific embodiment
With reference to the attached drawing in the embodiment of the present invention, technical solution in the embodiment of the present invention carries out clear, complete
Ground description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.Based on this
The embodiment of invention, every other implementation obtained by those of ordinary skill in the art without making creative efforts
Example, belongs to protection scope of the present invention.
Natural scene be usually it is complicated and changeable, for image object counting load, it is easy to the shadow by various factors
Ring, as between target seriously block, target deformation, the uneven distribution of target, mixed and disorderly background interference, camera angles it is abnormal
Become etc..The especially influence of video camera perspective effect, so that size variation multiplicity of the same object in scene different depth, different
The camera angles of scene equally change different.As previously mentioned, being directed to these problems, existing method mainly passes through introducing multi path network
Network extracts Analysis On Multi-scale Features, however, network parameter quantity can be greatly increased by introducing multi-channel network, improves computation complexity, nothing
Method meets application request.On the other hand, compared to single network model, the training of multi-channel network is usually extremely difficult
's.In fact, convolutional neural networks model, itself is a pyramid multi-level structure, model receives original image signal work
For input, layer-by-layer abstract expression is carried out to image, higher layer has bigger receptive field, contains between each layer feature rich
Rich multi-scale information.Therefore, the invention discloses a kind of image object method of counting based on convolutional neural networks, using list
A network carries out object count, and the multi-scale information for making full use of convolutional neural networks model itself to be included is reducing model
While complexity, good counting properties are achieved;Below for provided in an embodiment of the present invention a kind of based on convolutional Neural
The image object method of counting of network does detailed introduction.
As shown in Figure 1, a kind of image object method of counting based on convolutional neural networks is provided for the embodiment of the present invention,
In the 1st~4th step be the training stage, the 5th step is test phase;
Image used in training stage can come from representative crowd's enumeration data collection Shanghaitech-B
And the scene picture in vehicle count data set TRANCOS.Wherein Shanghaitech-B data set is by (Single-image
crowd counting via multi-column convolutional neural network.Proceedings of
IEEE Conference on Computer Vision and Pattern Recognition, 2016.) it provides, TRANCOS
Data set is by (Extremely Overlapping Vehicle Counting.Proceedings of Iberian
Conference on Pattern Recognition and Image Analysis, 2015.) it provides.
The key step of the above method is as follows:
1, pyramid object count network is established based on convolutional neural networks.
In the embodiment of the present invention, according to the characteristics of image object counting load itself and requirement, for representative convolution
Neural network model carries out structural adjustment and design, establishes pyramid object count network.As shown in Fig. 2, the gold established
Word tower object count network mainly includes:Characteristic extracting module, robustness enhancing module and density estimation module.
1) characteristic extracting module uses full convolutional network and extracts characteristics of image, including five layers of regular volume lamination and
Two layers of empty convolutional layer.
In the embodiment of the present invention, the characteristics of image under different scenes is extracted using two different network structures.Phase
Ying Di, using two kinds of various sizes of image blocks, 72 × 72 and 144 × 144, respectively as the first and second of network structure
Input size.The difference of two kinds of network structures is mainly that the receptive field of three first layers convolutional layer convolution kernel is of different sizes, the
A kind of network structure all uses the convolution kernel of 3 × 3 sizes, and the three first layers of second of network structure use the convolution of 5 × 5 sizes
Core.In two network structures, the port number of each layer is all since 16, and every by a maximum pond layer, port number increases by 2
Times, 16 then are reduced to from 64 again and are remained unchanged.Comprising two maximum pond layers in network, it is located at first layer convolutional layer
After second layer convolutional layer, core size is 2 × 2, step-length 1.
It should be noted that the above-mentioned size about image block, specific value used in number of channels are merely illustrative, and
It is non-to be construed as limiting.
Characteristic extracting module is finally two empty convolutional layers, and the cavity convolutional layer refers to the convolution kernel in Standard convolution
Middle injection cavity, the spacing being respectively worth when to increasing convolution kernel processing data can not increase compared to regular volume lamination
Expand the size of receptive field in the case where network parameter.
Illustratively, two empty convolutional layers can export every layer of convolution with step-length for 2, single using linear amendment
Nonlinear Mapping modeling ability is added as activation primitive, for network in first ReLU.
2) robustness enhances module, uses spatial pyramid pond mode, in the spy of characteristic extracting module output
It levies on figure, passes through N1×N1、N2×N2、N3×N3With N4×N4The space pond of four different scales, constructs various sizes of son
Block, to obtain spatial information of the image on different resolution, make e-learning to feature the deformation of target is had more
Robustness, while reducing the computation complexity of model.
Illustratively, the space pondization of four different scales can be 1 × 1,2 × 2,4 × 4 and 6 × 6, then binding characteristic
Example in extraction module, the characteristic dimension after robustness enhancing are 16 × (1 × 1+2 × 2+4 × 4+6 × 6)=912.
3) the density estimation module uses a kind of Pyramid technology counting module, learns on different scale complementary
Information, to generate target areal density figure.
The target density estimation that the Pyramid technology counting module is carried out is to pass through the enhanced feature of robustness
It is carried out respectively on each characteristic pattern of extraction module, the density map of final output is obtained by the output results added of each layer;Wherein, close
Degree estimation establishes the Nonlinear Mapping from characteristics of image to density value using two layers of full articulamentum.The full articulamentum of first layer and Shandong
The output feature of stick enhancing module (SPP) is connected, and neuron number is 1000.The full articulamentum of the second layer is output layer, when
When characteristic extracting module uses the first network structure, output neuron number is 324, when characteristic extracting module uses second
When network structure, output neuron number is 1296.After first layer connects entirely, using amendment linear unit (ReLU) activation
Function and Dropout layers, wherein the parameter of Dropout is 0.5.
It should be noted that it is above-mentioned about neuron number, Dropout parameter used in specific value be only show
Example, is not construed as limiting.
2, it is marked using artificial data, the Density Distribution true value image of interesting target is established on training image.
In the embodiment of the present invention, gaussian filtering is carried out using the target's center's point diagram manually marked on training image and is obtained
The Density Distribution true value image of interesting target;
Wherein, using the target's center position of target's center's point diagram of mark as the center of Gaussian kernel, pass through gaussian filtering
Generate density profile:As shown in figure 3, given training image, if P is the set of target geometric center point in the image of mark,
With D indicate image corresponding to density profile, then be located at (i, j) at pixel density value D (i, j) pass through following formula meter
It calculates:
In above formula,The dimensional gaussian distribution value of pixel at (i, j), Gaussian Profile it is equal
Value point is located at mark position (m, n);σ2I2×2For covariance matrix.
Illustratively, when it is implemented, for Shanghaitech-B and TRANCOS data set, the size of Gaussian kernel can
It is set to 10 and 15.
3, the training image and corresponding density point training data concentrated by random cropping and flip horizontal mode
Cloth true value image carries out enumeration data enhancing.
The parameter of convolutional neural networks model training is more, needs that a could be trained roll up based on a large amount of training data
Product neural network model.Therefore in the training stage, training data is enhanced by the method for the random cropping from training image, thus
Generate a large amount of training image blocks and corresponding real density figure.Dimension normalization is carried out according to input size, and random
It cuts out a large amount of image block and progress data enhancing is overturn by image level again, obtained training sample is finally used for model
Training.
Key step is as follows:
1) size of the training image of input is normalized;
2) from the training image after normalization the image block of random cropping same size as new training image;
3) flip horizontal is carried out to new training image, obtains a series of new training images;
4) aforesaid way (i.e. step 1)~3) are utilized) same treatment is done to Density Distribution true value image, then pass through normalizing
Change makes after scaling that destination number remains unchanged in Density Distribution true value image.
Illustratively, for Shanghaitech-B data set, according to document (Single-image crowd
counting via multi-column convolutional neural network.Proceedings of IEEE
Conference on Computer Vision and PatternRecognition, 2016.) experimental setup in, uses
For 400 pictures as training image, remaining 316 are test picture.The picture that Shanghaitech-B data set provides is differentiated
Rate is larger, and in order to train counter model, for every trained picture, it is 200 that our random croppings 200 when implementing, which are opened small greatly,
× 200 image block then by each image block scaling (normalizing) to 144 × 144, and uses second of feature extraction net
Network carries out feature extraction.Data enhancing is equally carried out by Image Reversal when implementation.Certainly, above-mentioned specific value is also only act
Example, is not construed as limiting.
4, using the enhanced training image of enumeration data and target density distribution true value image as pyramid object count
The input of network completes pyramid object count network training by continuous iteration optimization, generates pyramid object count network
Model.
When being trained to pyramid object count network, by the enhanced training image of enumeration data and target density point
Cloth true value image is as training sample, using the Euclidean distance between the density map and real density figure of prediction as loss letter
Number updates the model parameter of network, loss function L (Θ) by the training of stochastic gradient descent method in Optimized Iterative each time
It is defined as follows:
In above formula, Θ indicates the network parameter that model learning arrives, and N is training samples number, F (Xk;It Θ) is pyramid mesh
Mark the density map of counting and network prediction, DkIndicate k-th of training sample XkReal density figure.
As shown in Figure 4, it is determined that after Euclidean distance is as optimization aim, in the training of model, arrived first by end
The training at end, in the last layer feature (the last one empty convolutional layer of the characteristic extracting module after robustness enhancing
The characteristic pattern of output) on establish density regression model, obtain initial density estimation result.Then, in order to optimize count results,
Fixed character extraction module and initial regression model, using the residual error of real density and current predictive density as optimization aim,
In another layer of feature (feature of the last layer regular volume lamination output) of characteristic extracting module after enhancing by robustness
Establish a new regression model.Back-propagation algorithm is finally recycled to carry out joint training to the parameter of whole network.Pass through
This mode allows two regression models to learn complementary information from the Analysis On Multi-scale Features of convolutional neural networks, common to complete finally
Density estimation.According to this Training strategy, can be trained in Pyramid technology counting module using similar method more
Regression model.When it is implemented, final counting and network, which uses two regression models, carries out density estimations, two of use
Regression model is built upon respectively on the last layer cavity convolutional layer and the last layer regular volume lamination of feature extraction network.
5, when new images input, the image with input picture block same size is generated by sliding window mode, is sent to gold
In word tower object count network model, the density profile predicted is averaged to obtain final to the density value of lap
Output density figure, to acquire target numbers.
In order to illustrate the performance of above scheme of the present invention, also it is tested and assesses.
It, below will be to net after counting and network model has been respectively trained on Shanghaitech-B and TRANCOS data set
The performance of network is assessed, and the specific method is as follows:In test data set, for every test image, using 10 pixels as step-length
Sliding window is carried out, the image with input picture block same size is generated, is sent in trained counting and network model, is predicted
Density profile, finally the density value of lap is averaged to obtain final output density figure.By test data set
Real density figure is compared with predicted density figure, obtains assessment result.Fig. 5 and Fig. 6 gives prediction result schematic diagram.Fig. 5
In Fig. 6, first is classified as input picture, and second is classified as Density Distribution true value image, and third is classified as inventive network model prediction
Density map, the digital representation destination number below density map.
For Shanghaitech-B data set, using mean absolute error (MAE) and root-mean-square error (RMSE) conduct
Evaluation index, corresponding formula are as follows:
In above-mentioned formula, N indicates test sample quantity, CkFor the destination number for including in the kth picture of model prediction,It is corresponding authentic specimen quantity.
Table 1 be the present invention on Shanghaitech-B data set with the comparing result of existing method.It can be seen that this hair
It is bright that there is very high crowd to count accuracy rate.
1 comparing result of table
For TRANCOS data set, using net lattice control absolute error (Grid Average Mean Absolute
Error, GAME) it is used as evaluation index.GAME index considers the precision of counting and the accuracy to target distribution positioning simultaneously.
4 are divided by picture for specified scale L, GAME (L)LThen a Non-overlapping Domain calculates the average absolute in each region
Error, specific formula are as follows:
In above-mentioned formula, N indicates test sample quantity,For in the kth picture of model prediction first of region packet
The destination number contained,It is corresponding authentic specimen quantity.Particularly, GAME (0) and MAE evaluation criterion are of equal value.
Table 2 be the present invention on TRANCOS data set with the comparing result of existing method.As can be seen that the present invention has
Very high vehicle count accuracy rate.
2 comparing result of table
In above scheme of the embodiment of the present invention, for the problem that actual scene is complicated and changeable, have the advantage that:Firstly,
Enhance the feature that module arrives e-learning by robustness and robustness is had more to target deformation, while reducing the calculating of model
Complexity;Then, density estimation is carried out using Pyramid technology counting module, made full use of in convolutional neural networks layered characteristic
The multi-scale information for including significantly improves computational efficiency while realizing accurate counting.In short, the present invention is based on convolution minds
Through network, the target accurate metering in image is realized, can be adapted for the object count task under complex scene, is calculated complicated
Spend low, practical application value with higher.
Through the above description of the embodiments, those skilled in the art can be understood that above-described embodiment can
The mode of necessary general hardware platform can also be added to realize by software by software realization.Based on this understanding,
The technical solution of above-described embodiment can be embodied in the form of software products, which can store non-easy at one
In the property lost storage medium (can be CD-ROM, USB flash disk, mobile hard disk etc.), including some instructions are with so that a computer is set
Standby (can be personal computer, server or the network equipment etc.) executes method described in each embodiment of the present invention.
The foregoing is only a preferred embodiment of the present invention, but scope of protection of the present invention is not limited thereto,
Within the technical scope of the present disclosure, any changes or substitutions that can be easily thought of by anyone skilled in the art,
It should be covered by the protection scope of the present invention.Therefore, protection scope of the present invention should be with the protection model of claims
Subject to enclosing.
Claims (6)
1. a kind of image object method of counting based on convolutional neural networks, which is characterized in that including:
Pyramid object count network is established based on convolutional neural networks;
It is marked using artificial data, the Density Distribution true value image of interesting target is established on training image;
The training image that training data is concentrated by random cropping and flip horizontal mode and corresponding Density Distribution true value
Image carries out enumeration data enhancing;
Using the enhanced training image of enumeration data and target density distribution true value image as pyramid object count network
Input completes pyramid object count network training by continuous iteration optimization, generates pyramid object count network model;
When new images input, the image with input picture block same size is generated by sliding window mode, is sent to pyramid mesh
It marks in counting and network model, the density profile predicted is averaged the density value of lap to obtain final output
Density map, to acquire target numbers.
2. a kind of image object method of counting based on convolutional neural networks according to claim 1, which is characterized in that institute
The pyramid object count network of foundation includes:Characteristic extracting module, robustness enhancing module and density estimation module;Wherein:
The characteristic extracting module uses full convolutional network and extracts characteristics of image, including five layers of regular volume lamination and two layers of sky
Hole convolutional layer;The cavity convolutional layer refers to injects cavity in the convolution kernel of Standard convolution, to increase convolution kernel processing number
According to when the spacing that is respectively worth;
The robustness enhances module, uses spatial pyramid pond mode, on the characteristic pattern of characteristic extracting module output,
Pass through N1×N1、N2×N2、N3×N3With N4×N4The space pond of four different scales, constructs various sizes of sub-block, to obtain
Take spatial information of the image on different resolution;
The density estimation module uses a kind of Pyramid technology counting module, learns complementary information on different scale, from
And generate target areal density figure.
3. a kind of image object method of counting based on convolutional neural networks according to claim 2, which is characterized in that institute
Stating the target density estimation that Pyramid technology counting module is carried out is each by the enhanced characteristic extracting module of robustness
It is carried out respectively on layer characteristic pattern, the density map of final output is obtained by the output results added of each layer;Wherein, density estimation is adopted
The Nonlinear Mapping from characteristics of image to density value is established with two layers of full articulamentum.
4. a kind of image object method of counting based on convolutional neural networks according to claim 1, which is characterized in that institute
It states and is marked using artificial data, the Density Distribution true value image that interesting target is established on training image includes:
Gaussian filtering, which is carried out, using the target's center's point diagram manually marked on training image obtains the density point of interesting target
Cloth true value image;
Wherein, it using the target's center position of target's center's point diagram of mark as the center of Gaussian kernel, is generated by gaussian filtering
Density profile:If P be mark image in target geometric center point set, with D indicate image corresponding to Density Distribution
Figure, then the density value D (i, j) for being located at pixel at (i, j) are calculated by following formula:
In above formula,It is the dimensional gaussian distribution value of pixel at (i, j), the average point of Gaussian Profile
At mark position (m, n);σ2I2×2For covariance matrix.
5. a kind of image object method of counting based on convolutional neural networks according to claim 1, which is characterized in that institute
State the training image and corresponding Density Distribution true value figure concentrated by random cropping and flip horizontal mode to training data
As the step of carrying out enumeration data enhancing includes:
The size of the training image of input is normalized;
The image block of random cropping same size is as new training image from the training image after normalization;
Flip horizontal is carried out to new training image, obtains a series of new training images;
Same treatment is done to Density Distribution true value image using aforesaid way, then keeps Density Distribution after scaling true by normalization
Destination number remains unchanged in value image.
6. a kind of image object method of counting based on convolutional neural networks according to claim 1, which is characterized in that right
When pyramid object count network is trained, the enhanced training image of enumeration data and target density are distributed true value image
As training sample, using the Euclidean distance between the density map and real density figure of prediction as loss function, by random
Gradient descent method training, updates the model parameter of network in Optimized Iterative each time, and loss function L (Θ) is defined as follows:
In above formula, Θ indicates the network parameter that model learning arrives, and N is training samples number, F (Xk;It Θ) is pyramid object count
The density map of neural network forecast, DkIndicate k-th of training sample XkReal density figure.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810564162.6A CN108921822A (en) | 2018-06-04 | 2018-06-04 | Image object method of counting based on convolutional neural networks |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810564162.6A CN108921822A (en) | 2018-06-04 | 2018-06-04 | Image object method of counting based on convolutional neural networks |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108921822A true CN108921822A (en) | 2018-11-30 |
Family
ID=64419585
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810564162.6A Pending CN108921822A (en) | 2018-06-04 | 2018-06-04 | Image object method of counting based on convolutional neural networks |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108921822A (en) |
Cited By (26)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109584142A (en) * | 2018-12-05 | 2019-04-05 | 网易传媒科技(北京)有限公司 | Image Intensified System and method, training method, medium and electronic equipment |
CN109741301A (en) * | 2018-12-19 | 2019-05-10 | 北京理工大学 | A kind of intensive object count method based on deep learning faster |
CN109829923A (en) * | 2018-12-24 | 2019-05-31 | 五邑大学 | A kind of antenna for base station based on deep neural network has a down dip angle measuring system and method |
CN110009027A (en) * | 2019-03-28 | 2019-07-12 | 腾讯科技(深圳)有限公司 | Comparison method, device, storage medium and the electronic device of image |
CN110009611A (en) * | 2019-03-27 | 2019-07-12 | 中南民族大学 | A kind of sensation target dynamic itemset counting method and system towards image sequence |
CN110110666A (en) * | 2019-05-08 | 2019-08-09 | 北京字节跳动网络技术有限公司 | Object detection method and device |
CN110176023A (en) * | 2019-04-29 | 2019-08-27 | 同济大学 | A kind of light stream estimation method based on pyramid structure |
CN110263732A (en) * | 2019-06-24 | 2019-09-20 | 京东方科技集团股份有限公司 | Multiscale target detection method and device |
CN110705344A (en) * | 2019-08-21 | 2020-01-17 | 中山大学 | Crowd counting model based on deep learning and implementation method thereof |
CN110765833A (en) * | 2019-08-19 | 2020-02-07 | 中云智慧(北京)科技有限公司 | Crowd density estimation method based on deep learning |
CN110991317A (en) * | 2019-11-29 | 2020-04-10 | 中山大学 | Crowd counting method based on multi-scale perspective sensing type network |
CN110991225A (en) * | 2019-10-22 | 2020-04-10 | 同济大学 | Crowd counting and density estimation method and device based on multi-column convolutional neural network |
CN111242036A (en) * | 2020-01-14 | 2020-06-05 | 西安建筑科技大学 | Crowd counting method based on encoding-decoding structure multi-scale convolutional neural network |
CN111259833A (en) * | 2020-01-20 | 2020-06-09 | 青岛大学 | Vehicle counting method based on traffic images |
CN111429466A (en) * | 2020-03-19 | 2020-07-17 | 北京航空航天大学 | Space-based crowd counting and density estimation method based on multi-scale information fusion network |
CN111583655A (en) * | 2020-05-29 | 2020-08-25 | 苏州大学 | Traffic flow detection method, device, equipment and medium |
CN111709908A (en) * | 2020-05-09 | 2020-09-25 | 上海健康医学院 | Helium bubble segmentation counting method based on deep learning |
CN111738136A (en) * | 2020-06-19 | 2020-10-02 | 新希望六和股份有限公司 | Method and device for determining number of microscopic objects, computer equipment and storage medium |
CN111914765A (en) * | 2020-08-05 | 2020-11-10 | 杭州像素元科技有限公司 | Service area environment comfort level detection method and device and readable storage medium |
CN113011329A (en) * | 2021-03-19 | 2021-06-22 | 陕西科技大学 | Pyramid network based on multi-scale features and dense crowd counting method |
CN113284164A (en) * | 2021-05-19 | 2021-08-20 | 中国农业大学 | Shrimp swarm automatic counting method and device, electronic equipment and storage medium |
CN113343790A (en) * | 2021-05-21 | 2021-09-03 | 中车唐山机车车辆有限公司 | Traffic hub passenger flow statistical method, device and storage medium |
US11393182B2 (en) | 2020-05-29 | 2022-07-19 | X Development Llc | Data band selection using machine learning |
CN115619776A (en) * | 2022-12-02 | 2023-01-17 | 湖北凯乐仕通达科技有限公司 | Article counting method and device based on deep learning |
US11606507B1 (en) | 2020-08-28 | 2023-03-14 | X Development Llc | Automated lens adjustment for hyperspectral imaging |
US11651602B1 (en) | 2020-09-30 | 2023-05-16 | X Development Llc | Machine learning classification based on separate processing of multiple views |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106650913A (en) * | 2016-12-31 | 2017-05-10 | 中国科学技术大学 | Deep convolution neural network-based traffic flow density estimation method |
CN107301387A (en) * | 2017-06-16 | 2017-10-27 | 华南理工大学 | A kind of image Dense crowd method of counting based on deep learning |
CN107679477A (en) * | 2017-09-27 | 2018-02-09 | 深圳市未来媒体技术研究院 | Face depth and surface normal Forecasting Methodology based on empty convolutional neural networks |
CN107944327A (en) * | 2016-10-10 | 2018-04-20 | 杭州海康威视数字技术股份有限公司 | A kind of demographic method and device |
CN107967451A (en) * | 2017-11-23 | 2018-04-27 | 常州大学 | A kind of method for carrying out crowd's counting to static image using multiple dimensioned multitask convolutional neural networks |
TWI624805B (en) * | 2017-03-27 | 2018-05-21 | 晶睿通訊股份有限公司 | Object counting method having route distribution property and related image processing device |
-
2018
- 2018-06-04 CN CN201810564162.6A patent/CN108921822A/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107944327A (en) * | 2016-10-10 | 2018-04-20 | 杭州海康威视数字技术股份有限公司 | A kind of demographic method and device |
CN106650913A (en) * | 2016-12-31 | 2017-05-10 | 中国科学技术大学 | Deep convolution neural network-based traffic flow density estimation method |
TWI624805B (en) * | 2017-03-27 | 2018-05-21 | 晶睿通訊股份有限公司 | Object counting method having route distribution property and related image processing device |
CN107301387A (en) * | 2017-06-16 | 2017-10-27 | 华南理工大学 | A kind of image Dense crowd method of counting based on deep learning |
CN107679477A (en) * | 2017-09-27 | 2018-02-09 | 深圳市未来媒体技术研究院 | Face depth and surface normal Forecasting Methodology based on empty convolutional neural networks |
CN107967451A (en) * | 2017-11-23 | 2018-04-27 | 常州大学 | A kind of method for carrying out crowd's counting to static image using multiple dimensioned multitask convolutional neural networks |
Non-Patent Citations (4)
Title |
---|
SAIHUI HOU等: "DualNet: Learn Complementary Features for Image Recognition", 《CONFERENCE: 2017 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV)》 * |
VICTOR LEMPITSKY等: "Learning To Count Objects in Images", 《INFORMATION PROCESSING SYSTEMS 2010》 * |
YUHONG LI等: "CSRNet: Dilated Convolutional Neural Networks for Understanding the Highly Congested Scenes", 《ARXIV》 * |
时增林等: "基于序的空间金字塔池化网络的人群计数方法", 《自动化学报》 * |
Cited By (37)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109584142A (en) * | 2018-12-05 | 2019-04-05 | 网易传媒科技(北京)有限公司 | Image Intensified System and method, training method, medium and electronic equipment |
CN109741301A (en) * | 2018-12-19 | 2019-05-10 | 北京理工大学 | A kind of intensive object count method based on deep learning faster |
CN109829923A (en) * | 2018-12-24 | 2019-05-31 | 五邑大学 | A kind of antenna for base station based on deep neural network has a down dip angle measuring system and method |
CN110009611A (en) * | 2019-03-27 | 2019-07-12 | 中南民族大学 | A kind of sensation target dynamic itemset counting method and system towards image sequence |
CN110009611B (en) * | 2019-03-27 | 2021-05-14 | 中南民族大学 | Visual target dynamic counting method and system for image sequence |
CN110009027A (en) * | 2019-03-28 | 2019-07-12 | 腾讯科技(深圳)有限公司 | Comparison method, device, storage medium and the electronic device of image |
CN110009027B (en) * | 2019-03-28 | 2022-07-29 | 腾讯科技(深圳)有限公司 | Image comparison method and device, storage medium and electronic device |
CN110176023A (en) * | 2019-04-29 | 2019-08-27 | 同济大学 | A kind of light stream estimation method based on pyramid structure |
CN110176023B (en) * | 2019-04-29 | 2023-06-02 | 同济大学 | Optical flow estimation method based on pyramid structure |
CN110110666A (en) * | 2019-05-08 | 2019-08-09 | 北京字节跳动网络技术有限公司 | Object detection method and device |
CN110263732A (en) * | 2019-06-24 | 2019-09-20 | 京东方科技集团股份有限公司 | Multiscale target detection method and device |
CN110765833A (en) * | 2019-08-19 | 2020-02-07 | 中云智慧(北京)科技有限公司 | Crowd density estimation method based on deep learning |
CN110705344B (en) * | 2019-08-21 | 2023-03-28 | 中山大学 | Crowd counting model based on deep learning and implementation method thereof |
CN110705344A (en) * | 2019-08-21 | 2020-01-17 | 中山大学 | Crowd counting model based on deep learning and implementation method thereof |
CN110991225A (en) * | 2019-10-22 | 2020-04-10 | 同济大学 | Crowd counting and density estimation method and device based on multi-column convolutional neural network |
CN110991317A (en) * | 2019-11-29 | 2020-04-10 | 中山大学 | Crowd counting method based on multi-scale perspective sensing type network |
CN110991317B (en) * | 2019-11-29 | 2023-05-16 | 中山大学 | Crowd counting method based on multi-scale perspective sensing network |
CN111242036B (en) * | 2020-01-14 | 2023-05-09 | 西安建筑科技大学 | Crowd counting method based on multi-scale convolutional neural network of encoding-decoding structure |
CN111242036A (en) * | 2020-01-14 | 2020-06-05 | 西安建筑科技大学 | Crowd counting method based on encoding-decoding structure multi-scale convolutional neural network |
CN111259833A (en) * | 2020-01-20 | 2020-06-09 | 青岛大学 | Vehicle counting method based on traffic images |
CN111429466A (en) * | 2020-03-19 | 2020-07-17 | 北京航空航天大学 | Space-based crowd counting and density estimation method based on multi-scale information fusion network |
CN111709908A (en) * | 2020-05-09 | 2020-09-25 | 上海健康医学院 | Helium bubble segmentation counting method based on deep learning |
CN111709908B (en) * | 2020-05-09 | 2024-03-26 | 上海健康医学院 | Helium bubble segmentation counting method based on deep learning |
CN111583655A (en) * | 2020-05-29 | 2020-08-25 | 苏州大学 | Traffic flow detection method, device, equipment and medium |
CN111583655B (en) * | 2020-05-29 | 2021-12-24 | 苏州大学 | Traffic flow detection method, device, equipment and medium |
US11393182B2 (en) | 2020-05-29 | 2022-07-19 | X Development Llc | Data band selection using machine learning |
US11620804B2 (en) | 2020-05-29 | 2023-04-04 | X Development Llc | Data band selection using machine learning |
CN111738136A (en) * | 2020-06-19 | 2020-10-02 | 新希望六和股份有限公司 | Method and device for determining number of microscopic objects, computer equipment and storage medium |
CN111914765A (en) * | 2020-08-05 | 2020-11-10 | 杭州像素元科技有限公司 | Service area environment comfort level detection method and device and readable storage medium |
CN111914765B (en) * | 2020-08-05 | 2022-07-12 | 杭州像素元科技有限公司 | Service area environment comfort level detection method and device and readable storage medium |
US11606507B1 (en) | 2020-08-28 | 2023-03-14 | X Development Llc | Automated lens adjustment for hyperspectral imaging |
US11651602B1 (en) | 2020-09-30 | 2023-05-16 | X Development Llc | Machine learning classification based on separate processing of multiple views |
CN113011329B (en) * | 2021-03-19 | 2024-03-12 | 陕西科技大学 | Multi-scale feature pyramid network-based and dense crowd counting method |
CN113011329A (en) * | 2021-03-19 | 2021-06-22 | 陕西科技大学 | Pyramid network based on multi-scale features and dense crowd counting method |
CN113284164A (en) * | 2021-05-19 | 2021-08-20 | 中国农业大学 | Shrimp swarm automatic counting method and device, electronic equipment and storage medium |
CN113343790A (en) * | 2021-05-21 | 2021-09-03 | 中车唐山机车车辆有限公司 | Traffic hub passenger flow statistical method, device and storage medium |
CN115619776A (en) * | 2022-12-02 | 2023-01-17 | 湖北凯乐仕通达科技有限公司 | Article counting method and device based on deep learning |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108921822A (en) | Image object method of counting based on convolutional neural networks | |
CN106127204B (en) | A kind of multi-direction meter reading Region detection algorithms of full convolutional neural networks | |
CN108510012A (en) | A kind of target rapid detection method based on Analysis On Multi-scale Features figure | |
CN108171701B (en) | Significance detection method based on U network and counterstudy | |
Xue et al. | Remote sensing scene classification based on multi-structure deep features fusion | |
CN108986050A (en) | A kind of image and video enhancement method based on multiple-limb convolutional neural networks | |
CN108280233A (en) | A kind of VideoGIS data retrieval method based on deep learning | |
CN107967451A (en) | A kind of method for carrying out crowd's counting to static image using multiple dimensioned multitask convolutional neural networks | |
CN104992223A (en) | Dense population estimation method based on deep learning | |
CN106651830A (en) | Image quality test method based on parallel convolutional neural network | |
CN110263849A (en) | A kind of crowd density estimation method based on multiple dimensioned attention mechanism | |
CN107463919A (en) | A kind of method that human facial expression recognition is carried out based on depth 3D convolutional neural networks | |
CN107729819A (en) | A kind of face mask method based on sparse full convolutional neural networks | |
CN106920243A (en) | The ceramic material part method for sequence image segmentation of improved full convolutional neural networks | |
CN109165682A (en) | A kind of remote sensing images scene classification method merging depth characteristic and significant characteristics | |
CN108960404B (en) | Image-based crowd counting method and device | |
CN106991666B (en) | A kind of disease geo-radar image recognition methods suitable for more size pictorial informations | |
CN108491797A (en) | A kind of vehicle image precise search method based on big data | |
CN109948593A (en) | Based on the MCNN people counting method for combining global density feature | |
CN105678231A (en) | Pedestrian image detection method based on sparse coding and neural network | |
CN106960176B (en) | Pedestrian gender identification method based on transfinite learning machine and color feature fusion | |
CN109785344A (en) | The remote sensing image segmentation method of binary channel residual error network based on feature recalibration | |
CN111783589B (en) | Complex scene crowd counting method based on scene classification and multi-scale feature fusion | |
CN109543632A (en) | A kind of deep layer network pedestrian detection method based on the guidance of shallow-layer Fusion Features | |
CN107767416A (en) | The recognition methods of pedestrian's direction in a kind of low-resolution image |
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: 20181130 |
|
RJ01 | Rejection of invention patent application after publication |