CN107832835A - The light weight method and device of a kind of convolutional neural networks - Google Patents
The light weight method and device of a kind of convolutional neural networks Download PDFInfo
- Publication number
- CN107832835A CN107832835A CN201711123049.6A CN201711123049A CN107832835A CN 107832835 A CN107832835 A CN 107832835A CN 201711123049 A CN201711123049 A CN 201711123049A CN 107832835 A CN107832835 A CN 107832835A
- Authority
- CN
- China
- Prior art keywords
- matrix
- weight coefficient
- layer
- convolutional layer
- convolutional
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the present application discloses the light weight method and device of a kind of convolutional neural networks, and methods described includes：Obtain the weight coefficient matrix of each convolution kernel in convolutional neural networks model in every layer of convolutional layer；For any one weight coefficient matrix, the weight coefficient value that absolute value in the weight coefficient matrix is less than to predetermined threshold value is set to zero, obtains the first matrix and the second matrix；First matrix includes the weight coefficient of all non-zeros in the weight coefficient matrix, and second matrix includes subscript value of the weight coefficient of all non-zeros in the weight coefficient matrix in the weight coefficient matrix；Using first matrix and second matrix as the weight coefficient matrix after compression；The weight coefficient matrix after the compression of each convolution kernel in every layer of convolutional layer in the convolutional neural networks model carries out traffic congestion identification to the image of input.
Description
Technical field
The application is related to machine learning techniques field, the light weight method and dress of more particularly to a kind of convolutional neural networks
Put.
Background technology
In recent years, traffic congestion has become the bottleneck for restricting urban economy and social development, and it is urban road resource
For, need the concentrated reflection of contradiction, closely related with the various aspects of city operations, it directly contributes the overall operation efficiency in city
Low, the short -board effect during urban development is increasingly apparent.
The improvement of traffic congestion should be timely crossing, the section for finding congestion first, then utilize Traffic Announcement, microblogging etc.
New media public information platform sends congestion information, makes vehicle reasonable selection vehicle line, with reduce intersection, section pressure
Power, intersection, section is set progressively to recover unimpeded.
At present, when carrying out traffic congestion detection using video detection technology, there is two ways：One kind is transmitting video image
To the mode of Surveillance center；Another kind is to obtain the traffic such as flow, roadway occupancy, speed, following distance, queue length
After parameter, multiple traffic state datas therein are chosen, and realize to traffic congestion using pre-defined congestion method of discrimination
Judgement.Because the acquisition of various parameters is generally less accurate, so final process result can be caused to be forbidden, and which does not have
Preferable extended capability.Cause to calculate overlong time, have impact on the real-time of the identification of traffic congestion, and then can not timely lead to
Know vehicle reasonable selection vehicle line, it is little to the regulation effect of traffic congestion.
Therefore, the efficiency of road traffic image identification how is improved, is a urgent problem to be solved.
The content of the invention
The embodiment of the present application provides the light weight method and device of a kind of convolutional neural networks, is realized using deep learning
To traffic congestion crossing, the identification of the vehicle in section, and improve the efficiency of road traffic image identification.
The embodiment of the present application provides a kind of light weight method of convolutional neural networks, and methods described includes：
Obtain the weight coefficient matrix of each convolution kernel in convolutional neural networks model in every layer of convolutional layer；
For any one weight coefficient matrix, by weight system of the absolute value in the weight coefficient matrix less than predetermined threshold value
Numerical value is set to zero, obtains the first matrix and the second matrix；First matrix includes all non-in the weight coefficient matrix
Zero weight coefficient, second matrix include the weight coefficient of all non-zeros in the weight coefficient matrix in the weight coefficient
Subscript value in matrix；Using first matrix and second matrix as the weight coefficient matrix after compression；
The weight coefficient square after the compression of each convolution kernel in every layer of convolutional layer in the convolutional neural networks model
Battle array carries out traffic congestion identification to the image of input.
A kind of possible implementation, it is described acquisition the first matrix and the second matrix after, methods described also includes：
First matrix and second matrix are compressed using huffman coding.
A kind of possible implementation, the Section 1 of second matrix is all non-zeros in the weight coefficient matrix
The subscript value of the Section 1 of weight coefficient；The subscript value of the K items of second matrix is to own in the weight coefficient matrix
The difference of the subscript value of the subscript value of K item weight coefficients and K-1 items in the weight coefficient of non-zero, K are more than or equal to 2.
A kind of possible implementation, each volume in every layer of convolutional layer in the convolutional neural networks model
Weight coefficient matrix after product core compression carries out traffic congestion identification to the image of input, including：
Number of vehicles in described image and queue length are determined according to the convolutional neural networks model, and according to described
Number of vehicles and queue length in image determine the traffic congestion classification of described image.
A kind of possible implementation, the convolutional neural networks include M convolutional layer, and M is more than 2；
For the 2nd layer in the M convolutional layer to n-th of convolutional layer in M-1 layers, the convolutional layer includes the first son and rolled up
Lamination and the second sub- convolutional layer；The output channel of the first sub- convolutional layer connects with the input channel of the described second sub- convolutional layer
Connect；
The first sub- convolutional layer in the convolutional layer includes the convolution kernel of P 1 × 1；The P is the defeated of (n-1)th layer of convolutional layer
Go out port number；The second sub- convolutional layer in the convolutional layer includes the convolution kernel and the convolution kernel of L 3 × 3 of J 1 × 1, P=J+
L×3；The M, P, J, L are the positive integer more than or equal to 1, and the n is the positive integer more than 1 and less than M.
The embodiment of the present application provides a kind of lightweight device of convolutional neural networks, and described device includes：
Receiving unit, for obtaining the weight coefficient of each convolution kernel in convolutional neural networks model in every layer of convolutional layer
Matrix；
Processing unit, for for any one weight coefficient matrix, absolute value in the weight coefficient matrix being less than pre-
If the weight coefficient value of threshold value is set to zero, the first matrix and the second matrix are obtained；First matrix includes the weight system
The weight coefficient of all non-zeros in matrix number, second matrix include the weight system of all non-zeros in the weight coefficient matrix
Subscript value of the number in the weight coefficient matrix；Using first matrix and second matrix as the weight after compression
Coefficient matrix；The weight coefficient square after the compression of each convolution kernel in every layer of convolutional layer in the convolutional neural networks model
Battle array carries out traffic congestion identification to the image of input.
A kind of possible implementation, the processing unit compress first matrix and described the using huffman coding
Two matrixes.
A kind of possible implementation, the Section 1 of second matrix is all non-zeros in the weight coefficient matrix
The subscript value of the Section 1 of weight coefficient；The subscript value of the K items of second matrix is to own in the weight coefficient matrix
The difference of the subscript value of the subscript value of K item weight coefficients and K-1 items in the weight coefficient of non-zero, K are more than or equal to 2.
A kind of possible implementation, the processing unit are used for：
Number of vehicles in described image and queue length are determined according to the convolutional neural networks model, and according to described
Number of vehicles and queue length in image determine the traffic congestion classification of described image.
A kind of possible implementation, the convolutional neural networks include M convolutional layer, and M is more than 2；
For the 2nd layer in the M convolutional layer to n-th of convolutional layer in M-1 layers, the convolutional layer includes the first son and rolled up
Lamination and the second sub- convolutional layer；The output channel of the first sub- convolutional layer connects with the input channel of the described second sub- convolutional layer
Connect；
The first sub- convolutional layer in the convolutional layer includes the convolution kernel of P 1 × 1；The P is the defeated of (n-1)th layer of convolutional layer
Go out port number；The second sub- convolutional layer in the convolutional layer includes the convolution kernel and the convolution kernel of L 3 × 3 of J 1 × 1, P=J+
L×3；The M, P, J, L are the positive integer more than or equal to 1, and the n is the positive integer more than 1 and less than M.
The embodiment of the present application provides a kind of computer program product, including computer-readable instruction, when computer is read
And perform the computer-readable instruction so that computer performs the method as described in above-mentioned any one.
The embodiment of the present application provides a kind of chip, and the chip is connected with memory, for reading and performing the storage
The software program stored in device, to realize the method in any of the above-described in various possible designs.
The embodiment of the present application provides the light weight method and device of a kind of convolutional neural networks, by obtaining convolutional Neural
The weight system of each convolution kernel in network (Convolutional Neural Network, CNN) model in every layer of convolutional layer
Matrix number；For any one weight coefficient matrix, by weight system of the absolute value in the weight coefficient matrix less than predetermined threshold value
Numerical value is set to zero, obtains the first matrix and the second matrix；First matrix includes all non-in the weight coefficient matrix
Zero weight coefficient, second matrix include the weight coefficient of all non-zeros in the weight coefficient matrix in the weight coefficient
Subscript value in matrix；Using first matrix and second matrix as the weight coefficient matrix after compression；According to
The weight coefficient matrix after the compression of each convolution kernel in the convolutional neural networks model in every layer of convolutional layer is to the figure of input
As carrying out traffic congestion identification.On the premise of same identification precision is reached, to convolutional neural networks model of the prior art
Compression, reduce the calculating time of convolutional neural networks model, improve real-time and then the realization pair of the identification of traffic congestion
The judgement of traffic jam situation, congestion information is provided to driver in time, in order to shift to an earlier date programme path.
Brief description of the drawings
Fig. 1 is a kind of schematic flow sheet of the light weight method for convolutional neural networks that the embodiment of the present application provides；
Fig. 2 is a kind of structural representation for convolutional neural networks that the embodiment of the present application provides；
Fig. 3 is the structural representation for the convolutional neural networks that the embodiment of the present application provides；
Fig. 4 is a kind of structural representation of the lightweight device for convolutional neural networks that the embodiment of the present application provides.
Embodiment
Deep learning is a new field in machine learning research, and its motivation, which is to establish, simulates human brain is divided
Analyse the neutral net of study.It explains data, such as image, sound and text by imitating the mechanism of human brain, and its core is led to
Machine learning model and substantial amounts of training data of the structure with multiple hidden layers are crossed, to learn more useful feature, so as to final
Lifting classification or the accuracy of prediction.Convolutional neural networks are developed recentlies, and cause the one kind paid attention to extensively efficiently to be known
Other method.This method can carry out accurate judgement to the traffic conditions of present road, and provide the congestion level of traffic, be advantageous to
Traffic dispersion and supervision.The image comprising various traffic behaviors is learnt using deep learning method, obtains required net
Network model, then the road traffic picture currently collected can be handled automatically after model training is completed, to current
Traffic congestion situation is made decisions, and has preferable applicability and robustness, and reliable judgement is provided for road traffic supervision
Foundation.
The representative networks such as existing LeNet-5, AlexNet, GoogLeNet, ResNet are all based on convolutional Neural net
Network structure.Wherein, LeNet-5 is formed by 7 layers, and 1,3,5 layers are convolutional layer, and 2,4 layers are pond layer；AlexNet is by 8 layers of group
Into its first 5 layers are convolutional layers, and latter 3 layers are full articulamentums；GoogLeNet is formed by 22 layers, wherein there is 21 convolutional layers, 1 complete
Articulamentum；ResNet is formed by 152 layers, wherein there is 151 convolutional layers, last layer is full articulamentum.These network models are all
It is on the basis of convolutional neural networks model, by deepening improvement of the network depth completion to network model.In addition, somebody
Convolutional neural networks learning algorithm is combined by proposition with other thoughts or method, such as HMM and CNN model phases
With reference to mixed model, the model that quick PCA and CNN are combined etc..
In the prior art, it is smart in order to reach gratifying identification in the congestion identification model network based on deep learning
Degree, therefore all have selected using complicated network structure.In addition, it is above-mentioned based on neutral net in image classification method, on
The number and size of convolution kernel, pond method, activation primitive often rely on, are empirically determined, and inappropriate network structure can pole
The earth increase network parameter number, causes the increase of network calculations complexity.
When being detected using above-mentioned convolutional neural networks model of the prior art to road traffic image, model is relative
It is larger, cause to calculate overlong time, have impact on the real-time of the identification of traffic congestion, and then can not timely notify that vehicle is reasonable
Vehicle line is selected, it is little to the regulation effect of traffic congestion.
In addition, the communication between server at present is the limiting factor of the scalability of CNN training.For the parallel of data
Training, communication overhead are directly proportional to the number of parameters in model.Therefore, mini Mod is trained faster, since it is desired that less communication；
Secondly, expense is reduced when new model is exported into client, less model needs less communication, frequent updating more may be used
OK；
Current field programmable gate array (Field-Programmable Gate Array, FPGA) and embedded portion
Administration, FPG are typically below 10MB on-chip memory and no chip external memory.Sufficiently small model can be stored directly
On FPGA, so that FPGA can handle video flowing in real time.
As shown in figure 1, the embodiment of the present application provides a kind of light weight method schematic flow sheet of convolutional neural networks.
In flow shown in Fig. 1, on the premise of same identification precision is reached, to convolutional neural networks of the prior art
Model compression, to improve the real-time of object detection, improve the real-time of detection and the efficiency of detection.Comprise the following steps：
Step 101：Obtain the weight coefficient matrix of each convolution kernel in convolutional neural networks model in every layer of convolutional layer；
Step 102：For any one weight coefficient matrix, absolute value in the weight coefficient matrix is less than predetermined threshold value
Weight coefficient value be set to zero, obtain the first matrix and the second matrix；First matrix includes the weight coefficient matrix
In all non-zeros weight coefficient, second matrix includes the weight coefficient of all non-zeros in the weight coefficient matrix at this
Subscript value in weight coefficient matrix；Using first matrix and second matrix as the weight coefficient square after compression
Battle array；
Step 103：The power after the compression of each convolution kernel in every layer of convolutional layer in the convolutional neural networks model
Weight coefficient matrix carries out traffic congestion identification to the image of input.
The database of traffic image composition of the image of the embodiment of the present application from shooting, can be with the embodiment of the present application
The parts of images in the database is chosen as training image, parts of images makees test image, and training image and test image
It is not overlapping.In the embodiment of the present application, inputted by image to before convolutional neural networks, including：
Obtain training sample image or test sample image and it is pre-processed.
Specifically, training image and test image can be chosen from image data base, and training sample image is carried out
Pretreatment.Training image and test image can be simply pre-processed, are then saved in database.It is being embodied
During, the quantity of training data may be very big, and the efficiency for reading data from image file and being initialized is very
Low, so data are pre-stored in database, the rhythm of training can be accelerated.The pretreatment can subtract including image
Average, to improve training speed and precision.After carrying out image and subtracting average, the pixel of the training image or test image is entered
Row initialization process, to obtain the image to be entered of default pixel size.For example, it is 96 that first layer convolutional layer, which is pixel size,
× 96 training sample image.
As shown in Fig. 2 the embodiment of the present application, there is provided a kind of convolutional neural networks structural representation.
The convolutional neural networks include M layers convolutional layer and R layers pond layer, and M is more than 2；In M layer convolutional layers, first layer
It is non-full articulamentum to M-1 layers, last layer is full connection convolutional layer.First layer is convolutional layer, is inputted to first layer convolution
The image of layer is pretreated training image or test image.Output layer is connected after full articulamentum.
First layer is one layer of convolutional layer, extracts the training image or the T dimensional features of test image, as convolution god
T convolution kernel of the first layer through network model.It is exported as T characteristic pattern after T convolution kernel convolution.For the M layers
In convolutional layer the 2nd layer to the n-th layer convolutional layer (for convenience of description, hereinafter referred to as optimizing convolutional layer) in M-1 layers, optimization volume
Lamination includes the first sub- convolutional layer and the second sub- convolutional layer；The output channel of the first sub- convolutional layer and the described second sub- convolution
The input channel connection of layer；The n is the positive integer more than 1 and less than M, and the T, M are the positive integer more than 1.
The convolution kernel of the first sub- convolutional layer is convolution kernel corresponding to the output channel of preceding layer convolutional layer；Described second
Sub- convolutional layer includes first kind convolution kernel and the second class convolution kernel；The first kind convolution kernel for the first sub- convolutional layer output
P convolution kernel in convolution kernel number corresponding to passage；The P is the output channel number of (n-1)th layer of convolutional layer；
The size of the second class convolution kernel is A × A times of the convolution kernel size in first kind convolutional layer；Second class
The number of convolution kernel is J, and meets P=J+L × A；The P, J, L, A are the positive integer more than or equal to 1.
For example, for the 2nd layer in the M layers convolutional layer to the n-th layer convolutional layer in M-1 layers, in the convolutional layer
One sub- convolutional layer includes the convolution kernel of P 1 × 1；The P is the output channel number of (n-1)th layer of convolutional layer；In the convolutional layer
Two sub- convolutional layers include the convolution kernel and the convolution kernel of L 3 × 3 of J 1 × 1, P=J+L × 3；The P, J, L to be more than or
Positive integer equal to 1.
For example, if it is 4 to choose A, the number of the convolution kernel of the first sub- convolutional layer is 96, and pixel size is 1 × 1,
Then the number of the first kind convolution kernel of the second sub- convolutional layer is 24, and pixel size is 1 × 1, and the number of the second class convolution kernel is
24, pixel size is 3 × 3.Number is 48 after the characteristic pattern of the output of first layer optimization convolutional layer, and including 2 classes, first
The number of class convolution kernel is the 1/4 of the number of previous convolutional layer, i.e., 24, the convolution nuclear phase of size and the first sub- convolutional layer
Together, pixel size is 1 × 1；The number of second class convolution kernel for previous convolutional layer number 1/4, i.e., 24, pixel size
For 3 × 3.
The characteristic pattern of the input of every layer of optimization convolutional layer in the M-2 layers optimization convolutional layer optimizes convolutional layer for preceding layer
The characteristic pattern of output.The structure of every layer of optimization convolutional layer is essentially identical, and difference is the characteristic pattern number of every layer of optimization convolutional layer
With characteristic pattern corresponding to convolution kernel pixel size it is different.
For example, so that A is 4 as an example, the number of the convolution kernel of the first sub- convolutional layer is 48, and pixel size is 1 × 1；
The number of the first kind convolution kernel of second sub- convolutional layer is 12, and pixel size is 1 × 1, and the number of the second class convolution kernel is 12
Individual, pixel size is 3 × 3.
In addition, ReLU activation primitives are introduced after every layer of convolutional layer carries out convolution, it is every to obtain by ReLU activation primitives
The output of layer convolutional layer.Every layer of convolutional layer herein includes first layer convolutional layer, every layer of first sub- convolution optimized in convolutional layer
Layer and the second sub- convolutional layer, and the full articulamentum of last layer.
In CNN models in positive transmittance process, according to each optimization convolutional layer convolution kernel to described image convolution, obtain
The characteristic information of each optimization convolutional layer；Carried out obtained characteristic information as the output result of the current optimization convolutional layer
Output, using the input as next optimization convolutional layer；
After feature is obtained by convolution, it can be classified by these features.The feature of diverse location is entered
Row aggregate statistics, the feature after these aggregate statistics is compared not only has much lower dimension using all obtained features extracted
Degree, while can also improve result, it is not easy to over-fitting.This aggregate statistics are pond.
In specific implementation process, certain several layers of optimization during convolutional layer can be optimized with the 2nd to M-1 layers after the 1st layer
After convolutional layer, pond layer is added, reduces the Pixel Dimensions of the characteristic image of half.Pond can be determined according to the result of convolution
Change method.Pond layer in convolutional neural networks can select maximum pond method to carry out pond, can also select random pool
Method or average pond method carry out pond.Preferably, can be in first layer convolutional layer, third layer optimization convolutional layer, layer 7
Optimize convolutional layer and add pond layer afterwards.
In the training process, repeatedly down-sampled operation is carried out to sample image, the volume of the image is obtained after each convolution
Product characteristic pattern, pond layer carry out down-sampled operation to the convolution characteristic pattern, reach the purpose for reducing image resolution ratio, so as to reduce
Amount of calculation, improve the efficiency of detection.
Step 3: pretreated training image or test image are instructed according to the convolutional neural networks model of determination
Practice.
High-precision mode is put forward relative to the simple increase convolution number of plies in the prior art, by using the convolutional Neural net
The structure of network model, while the precision of convolutional neural networks model is improved, the convolution number of plies is limited, reduces parameter
Quantity, the substantial amounts of calculating time is saved, and then realized the lightweight to convolutional neural networks model.
CNN model training processes be divided into before to transmit and back transfer, the effect of forward direction transmission be to utilize convolutional Neural net
Network model is exported, and is obtained export structure, can be obtained levels of traffic congestion, and this result is considered as prediction category result, and
The composition of back transfer is that the parameter of the convolutional neural networks model is adjusted so that finally determines that a precision reaches pre-
If the convolutional neural networks model of condition.
Wherein, in forward direction transmittance process, the image of input is sent into the convolutional neural networks model, by preceding to biography
Pass the classification for determining the output of convolutional neural networks model.Its detailed process is：
Step 1: being inputted the image of the input as input information to the first sub- convolutional layer, input information is passed through
The weight coefficient matrix and bias term of every layer of the convolution kernel determined carry out convolution to the characteristic pattern of every layer of input, will pass through convolution
Core adds the information after bias term to obtain the characteristic pattern of each layer of output by a ReLU activation primitive.Increase bias term may be such that
Characteristics of image strengthens and can suppress noise.
In step 1, the weight coefficient matrix of the convolution kernel can step 102 acquisition according to.Therefore, right
, can be with the weight coefficient matrix after being compressed in obtaining step 102, described in decompression before the characteristic pattern of every layer of input carries out convolution
Weight coefficient matrix after compression.By the weight coefficient matrix after decompression, convolution is carried out with the value of the characteristic pattern of every layer of input.
By using the weight coefficient matrix after compression, the calculating of zero in weight coefficient matrix can be significantly reduced,
The substantial amounts of calculating time has been saved, has realized the lightweight of convolutional neural networks, has improved the real-time of the identification of traffic congestion
And then the timely judgement to traffic jam situation is realized, improve the efficiency of road traffic image identification.
Step 2: it regard the characteristic pattern of output as next layer of input；If next layer is convolutional layer, repeat step one,
If next layer is pond layer, into step 3；
Step 3: the 1/2 down-sampled down-sampled result for obtaining pixel and halving is done to the characteristic pattern of input；
Step 4: the output of last layer of non-full articulamentum is passed to full articulamentum, the characteristic pattern to inputting full articulamentum
In obtained characteristic vector, determine a classification score value for each pixel.By being integrated to obtained characteristic vector, obtain
To a subscript range vectors, every a kind of score value of test chart finally is obtained with a full figure, judges the traffic of the image of input
Congestion classification.Every a kind of score value can be probable value using softmax function normalizations.In forward-propagating process, i.e.,
Complete the identification to the classification of sample image, it can be determined that go out the levels of traffic congestion of input sample image.
For back transfer process, it is mainly used in before calculating to the classification and the loss letter of sample concrete class for transmitting output
Numerical value, loss function value is adjusted to the weight coefficient matrix of every layer of convolutional layer by the method back transfer of minimization error, obtained
Final convolutional neural networks model.
Specifically, weight coefficient corresponding to the characteristic pattern in every layer of convolutional layer can be determined by K-Means clustering methods
Matrix；Each corresponding convolution kernel of characteristic pattern, the corresponding weight coefficient matrix of each convolution kernel.Clustered according to K-Means
Method, the gradient for being attributed to of a sort node are added, and are taken in the M once resulting cluster that wherein target function value is minimum
The heart, the initialization value as the convolution kernel of the convolutional layer containing M characteristic pattern.
A kind of possible implementation, the weight coefficient matrix of the convolution kernel can be compressed according to step 102.Cause
This, in the characteristic pattern loss function value to every layer of input, loss function value is adjusted by the method back transfer of minimization error
After the weight coefficient matrix of every layer of convolutional layer, the weight coefficient matrix can be compressed by the method for step 102, and will
Weight coefficient matrix after the compression is stored.
Further to compress the weight coefficient matrix, a kind of possible implementation, the Section 1 of second matrix
For the subscript value of the Section 1 of the weight coefficient of all non-zeros in the weight coefficient matrix；The K items of second matrix
Subscript value is the subscript value and K-1 items of K item weight coefficients in the weight coefficient of all non-zeros in the weight coefficient matrix
Subscript value difference, K be more than or equal to 2.
Specifically, compress mode in a step 102 can be in the following manner：
Step 1: according to the weight coefficient matrix of acquisition, absolute value in the weight coefficient matrix is less than predetermined threshold value
Weight coefficient value is set to zero, obtains the first matrix and the second matrix；The predetermined threshold value can be less than 0.1.
For example, the weight coefficient matrix is following form：
| 0.08 | 0.92 | 0.52 |
| 0.07 | 0.41 | 0.78 |
| 0.29 | -0.04 | -0.34 |
| 0.54 | 0.41 | -0.45 |
The weight coefficient value that absolute value in the weight coefficient matrix is less than to predetermined threshold value according to step 1 is set to zero, obtains
The first matrix obtained can be following form：
| 0.92 | 0.52 | |
| 0.41 | 0.78 | |
| 0.29 | -0.34 | |
| 0.54 | 0.41 | -0.45 |
First matrix can be [0.92,0.52,0.41,0.78,0.29, -0.34,0.54,0.41, -0.45]；
Second matrix includes the weight coefficient of all non-zeros in the weight coefficient matrix in the weight coefficient matrix
In subscript value；For example, second matrix can be [1,2,4,5,6,8,9,10,11].
Further to compress the weight coefficient matrix, second matrix can be expressed as following form：Described second
The Section 1 of matrix is the subscript value of the Section 1 of the weight coefficient of all non-zeros in the weight coefficient matrix；Second square
The value of other each single item of battle array is the subscript value of the weight coefficient of all non-zeros and previous item in the weight coefficient matrix
Subscript value difference.For example, second matrix can be expressed as [1,1,2,1,1,2,1,1,1].
By using the weight coefficient matrix after compression, the calculating of zero in weight coefficient matrix can be significantly reduced,
The substantial amounts of calculating time has been saved, has realized the lightweight of convolutional neural networks, has improved the real-time of the identification of traffic congestion
And then the timely judgement to traffic jam situation is realized, improve the efficiency of road traffic image identification.
Further to realize the lightweight of convolutional neural networks, a kind of possible implementation, obtain the first matrix and
After second matrix, methods described also includes：First matrix and second matrix are compressed using huffman coding.Hough
The specific implementation of graceful coding, will not be repeated here.
In the embodiment of the present application, the training of the convolutional neural networks and testing procedure include：
Step 1: by pretreated training sample image input picture classify convolutional neural networks structure, by it is preceding to
Transmit and two step iterative cycles of back transfer update the weight coefficient matrix of CNN models, until reaching preparatory condition, train
Process terminates, the convolutional neural networks trained.
Untill the preparatory condition can be the error convergence between the congestion classification of prediction and the congestion classification of reality.
Tested, tested Step 2: pretreated test image to be input to the convolutional neural networks trained
Convolutional neural networks structure after card.
In the embodiment of the present application, each convolution kernel in every layer of convolutional layer in the convolutional neural networks model
Weight coefficient matrix after compression carries out traffic congestion identification to the image of input, including：
The vehicle characteristic information and link characteristic information determined according to the convolutional neural networks model, and according to the figure
Number of vehicles and queue length as in determine the traffic congestion classification of described image.
In specific implementation process, the convolutional neural networks model that can complete the incoming training of image in section is (most
Whole CNN models) in, to transmission (process of forward direction transmission herein and foregoing forward direction transmittance process one before carrying out
Cause), the output of CNN models is obtained, congestion classification corresponding to the road in picture is judged according to output result.Differentiate to improve
Accuracy, the result of multiframe can be merged as final traffic congestion classification.
Specifically, the number of vehicles and queuing length in described image can be determined according to the convolutional neural networks model
Degree, determine the traffic congestion classification of described image.
In specific implementation process, it can choose and be set up at least 1 in the camera in intersection section or common section
In the final convolutional neural networks model that the incoming training of two field picture is completed, to transmission before carrying out, final convolutional neural networks are obtained
The output of model, congestion classification corresponding to the road in current training image or test image is judged according to output result.According to
China《Road traffic blockage percentage and evaluation method (national standard)》Description for urban transportation traffic status is mainly in terms of two
To evaluate, i.e., intersection obstruction and section are blocked.Wherein intersection obstruction is defined as vehicle driveway outside intersection
Be obstructed queue length more than 500m for obstruction, 800m is Severe blockage；Section obstruction deliberated index is that length is more than 2000m
Obstruction, 3000m is Severe blockage.Traffic congestion classification in the embodiment of the present application can determine according to above deliberated index.Example
Such as, when section is common section, levels of traffic congestion is divided according to the standard in common section, when section is intersection
During section, levels of traffic congestion divides etc. according to the standard in intersection section.
The embodiment of the present application provides a kind of structure of convolutional neural networks model, the convolutional neural networks include convolutional layer,
Optimize convolutional layer, pond layer, full articulamentum and return layer.Wherein convolutional layer can use several convolution kernels to the image of input
Carry out convolution operation；Image progress of the pond layer to input is down-sampled, and a picture is merged into per horizontal and vertical two pixels
Vegetarian refreshments, reach the purpose for lowering image resolution ratio；Full articulamentum connects for common neutral net, each section of full articulamentum
Point is connected with the output node of last layer；Return layer and carry out back transfer process to the feature of input, cycle calculations, finally
Export recognition result.
The concrete structure of the convolutional neural networks in the image to driver as shown in figure 3, enter in the embodiment of the present application
After row normalization, image size is 96 × 96, and the size of the convolution kernel of first layer convolutional layer is 1 × 1.
As shown in figure 3, the image after normalization is input to convolutional neural networks, the convolutional layer of 11 × 1 is first passed through, then pass through
Cross a pond layer, the input of the output result of pond layer as one layer of optimization convolutional layer, then after 2 layers optimize convolutional layer,
Pass through a pond layer again, the inputs of the output result of pond layer as 5 layers of optimization convolutional layer, then by a pond layer, pond
Change layer to be connected with full articulamentum, returning layer by softmax is identified result.
In the embodiment of the present application, including substantial amounts of training image, corresponding detection can be used in each training image
Model, the wherein detection model include vehicle detection model and Road Detection model, according to the image information of each vehicle, to volume
Product neutral net is trained, according to the position where each vehicle, vehicle number in the road, queue length, it is determined that often
The jam situation of individual road.
It is normalized, returns before being trained to convolutional neural networks, in addition to the image of each input
The image size inputted after one change is identical.
As shown in figure 4, the embodiment of the present application provides a kind of lightweight device of convolutional neural networks, described device bag
Include：
Receiving unit 401, for obtaining the weight of each convolution kernel in convolutional neural networks model in every layer of convolutional layer
Coefficient matrix；
Processing unit 402, for for any one weight coefficient matrix, absolute value in the weight coefficient matrix to be less than
The weight coefficient value of predetermined threshold value is set to zero, obtains the first matrix and the second matrix；First matrix includes the weight
The weight coefficient of all non-zeros in coefficient matrix, second matrix include the weight of all non-zeros in the weight coefficient matrix
Subscript value of the coefficient in the weight coefficient matrix；Using first matrix and second matrix as the power after compression
Weight coefficient matrix；The weight coefficient after the compression of each convolution kernel in every layer of convolutional layer in the convolutional neural networks model
Matrix carries out traffic congestion identification to the image of input.
A kind of possible implementation, processing unit 402 compress first matrix and described the using huffman coding
Two matrixes.
A kind of possible implementation, the Section 1 of second matrix is all non-zeros in the weight coefficient matrix
The subscript value of the Section 1 of weight coefficient；The subscript value of the K items of second matrix is to own in the weight coefficient matrix
The difference of the subscript value of the subscript value of K item weight coefficients and K-1 items in the weight coefficient of non-zero, K are more than or equal to 2.
A kind of possible implementation, processing unit 402 are used for：
Number of vehicles in described image and queue length are determined according to the convolutional neural networks model, and according to described
Number of vehicles and queue length in image determine the traffic congestion classification of described image.
A kind of possible implementation, the convolutional neural networks include M convolutional layer, and M is more than 2；
For the 2nd layer in the M convolutional layer to n-th of convolutional layer in M-1 layers, the convolutional layer includes the first son and rolled up
Lamination and the second sub- convolutional layer；The output channel of the first sub- convolutional layer connects with the input channel of the described second sub- convolutional layer
Connect；
The first sub- convolutional layer in the convolutional layer includes the convolution kernel of P 1 × 1；The P is the defeated of (n-1)th layer of convolutional layer
Go out port number；The second sub- convolutional layer in the convolutional layer includes the convolution kernel and the convolution kernel of L 3 × 3 of J 1 × 1, P=J+
L×3；The M, P, J, L are the positive integer more than or equal to 1, and the n is the positive integer more than 1 and less than M.
The embodiment of the present application provides a kind of computer program product, including computer-readable instruction, when computer is read
And perform the computer-readable instruction so that computer performs the method as described in above-mentioned any one.
The embodiment of the present application provides a kind of chip, and the chip is connected with memory, for reading and performing the storage
The software program stored in device, to realize the method in any of the above-described in various possible designs.
Convolutional neural networks training method in the embodiment of the present application and the object identifying method based on convolutional neural networks,
It can implement in the various equipment for performing digital media signal processing, including：Computer, image and videograph, transmission and
Receiving device, portable video player, video conference etc..Above-mentioned technology can be implemented in hardware circuit, it is also possible to calculate
The digital media processing software that is performed in machine or other computing environment is realized.
The embodiment of the present application provides object detecting method and device in a kind of image, and convolutional neural networks have powerful
Feature learning ability, can overcome and describe the problem of not accurate enough as caused by artificial top set feature, can ensure it is accurate
On the basis of true rate, lower amount of calculation.The embodiment of the present application is believed using convolutional neural networks one-off recognition traffic congestion classification
Breath, it is easy to carry out global optimization to the detection process, and the weight coefficient matrix is have compressed when being detected, so as to drops
The low size of model, improves the real-time of detection and the efficiency of detection.
For systems/devices embodiment, because it is substantially similar to embodiment of the method, so the comparison of description is simple
Single, the relevent part can refer to the partial explaination of embodiments of method.
It should be understood by those skilled in the art that, embodiments herein can be provided as method, system or computer program
Product.Therefore, the application can use the reality in terms of complete hardware embodiment, complete software embodiment or combination software and hardware
Apply the form of example.Moreover, the application can use the computer for wherein including computer usable program code in one or more
The computer program production that usable storage medium is implemented on (including but is not limited to magnetic disk storage, CD-ROM, optical memory etc.)
The form of product.
The application is with reference to the flow according to the method for the embodiment of the present application, equipment (system) and computer program product
Figure and/or block diagram describe.It should be understood that can be by every first-class in computer program instructions implementation process figure and/or block diagram
Journey and/or the flow in square frame and flow chart and/or block diagram and/or the combination of square frame.These computer programs can be provided
The processors of all-purpose computer, special-purpose computer, Embedded Processor or other programmable data processing devices is instructed to produce
A raw machine so that produced by the instruction of computer or the computing device of other programmable data processing devices for real
The device for the function of being specified in present one flow of flow chart or one square frame of multiple flows and/or block diagram or multiple square frames.
These computer program instructions, which may be alternatively stored in, can guide computer or other programmable data processing devices with spy
Determine in the computer-readable memory that mode works so that the instruction being stored in the computer-readable memory, which produces, to be included referring to
Make the manufacture of device, the command device realize in one flow of flow chart or multiple flows and/or one square frame of block diagram or
The function of being specified in multiple square frames.
These computer program instructions can be also loaded into computer or other programmable data processing devices so that counted
Series of operation steps is performed on calculation machine or other programmable devices to produce computer implemented processing, so as in computer or
The instruction performed on other programmable devices is provided for realizing in one flow of flow chart or multiple flows and/or block diagram one
The step of function of being specified in individual square frame or multiple square frames.
Although having been described for the preferred embodiment of the application, those skilled in the art once know basic creation
Property concept, then can make other change and modification to these embodiments.So appended claims be intended to be construed to include it is excellent
Select embodiment and fall into having altered and changing for the application scope.
Obviously, those skilled in the art can carry out the essence of various changes and modification without departing from the application to the application
God and scope.So, if these modifications and variations of the application belong to the scope of the application claim and its equivalent technologies
Within, then the application is also intended to comprising including these changes and modification.
Claims (10)
1. a kind of light weight method of convolutional neural networks, it is characterised in that methods described includes：
Obtain the weight coefficient matrix of each convolution kernel in convolutional neural networks model in every layer of convolutional layer；
For any one weight coefficient matrix, by weight coefficient value of the absolute value in the weight coefficient matrix less than predetermined threshold value
Zero is set to, obtains the first matrix and the second matrix；First matrix includes all non-zeros in the weight coefficient matrix
Weight coefficient, second matrix include the weight coefficient of all non-zeros in the weight coefficient matrix in the weight coefficient matrix
In subscript value；Using first matrix and second matrix as the weight coefficient matrix after compression；
The weight coefficient matrix pair after the compression of each convolution kernel in every layer of convolutional layer in the convolutional neural networks model
The image of input carries out traffic congestion identification.
2. according to the method for claim 1, it is characterised in that after the first matrix of the acquisition and the second matrix, institute
Stating method also includes：
First matrix and second matrix are compressed using huffman coding.
3. according to the method for claim 1, it is characterised in that the Section 1 of second matrix is the weight coefficient square
The subscript value of the Section 1 of the weight coefficient of all non-zeros in battle array；The subscript value of the K items of second matrix is the weight
In coefficient matrix in the weight coefficient of all non-zeros the subscript value of the subscript value of K item weight coefficients and K-1 items difference, K
More than or equal to 2.
4. the method as described in claim 1, it is characterised in that described according to every layer of convolution in the convolutional neural networks model
The weight coefficient matrix after the compression of each convolution kernel in layer carries out traffic congestion identification to the image of input, including：
Number of vehicles in described image and queue length are determined according to the convolutional neural networks model, and according to described image
In number of vehicles and queue length determine the traffic congestion classification of described image.
5. according to the method for claim 1, it is characterised in that the convolutional neural networks include M convolutional layer, and M is more than
2；
For the 2nd layer in the M convolutional layer to n-th of convolutional layer in M-1 layers, the convolutional layer includes the first sub- convolutional layer
With the second sub- convolutional layer；The output channel of the first sub- convolutional layer is connected with the input channel of the described second sub- convolutional layer；
The first sub- convolutional layer in the convolutional layer includes the convolution kernel of P 1 × 1；The P is that the output of (n-1)th layer of convolutional layer is led to
Road number；Convolution kernel and the convolution kernel of L 3 × 3 of the second sub- convolutional layer including J individual 1 × 1 in the convolutional layer, P=J+L ×
3；The M, P, J, L are the positive integer more than or equal to 1, and the n is the positive integer more than 1 and less than M.
6. the lightweight device of a kind of convolutional neural networks, it is characterised in that described device includes：
Receiving unit, for obtaining the weight coefficient square of each convolution kernel in convolutional neural networks model in every layer of convolutional layer
Battle array；
Processing unit, for for any one weight coefficient matrix, absolute value in the weight coefficient matrix to be less than into default threshold
The weight coefficient value of value is set to zero, obtains the first matrix and the second matrix；First matrix includes the weight coefficient square
The weight coefficient of all non-zeros, the weight coefficient that second matrix includes all non-zeros in the weight coefficient matrix exist in battle array
Subscript value in the weight coefficient matrix；Using first matrix and second matrix as the weight coefficient after compression
Matrix；The weight coefficient matrix pair after the compression of each convolution kernel in every layer of convolutional layer in the convolutional neural networks model
The image of input carries out traffic congestion identification.
7. device according to claim 1, it is characterised in that the processing unit is using huffman coding compression described the
One matrix and second matrix.
8. device according to claim 1, it is characterised in that the Section 1 of second matrix is the weight coefficient square
The subscript value of the Section 1 of the weight coefficient of all non-zeros in battle array；The subscript value of the K items of second matrix is the weight
In coefficient matrix in the weight coefficient of all non-zeros the subscript value of the subscript value of K item weight coefficients and K-1 items difference, K
More than or equal to 2.
9. device as claimed in claim 1, it is characterised in that the processing unit is used for：
Number of vehicles in described image and queue length are determined according to the convolutional neural networks model, and according to described image
In number of vehicles and queue length determine the traffic congestion classification of described image.
10. device according to claim 1, it is characterised in that the convolutional neural networks include M convolutional layer, and M is more than
2；
For the 2nd layer in the M convolutional layer to n-th of convolutional layer in M-1 layers, the convolutional layer includes the first sub- convolutional layer
With the second sub- convolutional layer；The output channel of the first sub- convolutional layer is connected with the input channel of the described second sub- convolutional layer；
The first sub- convolutional layer in the convolutional layer includes the convolution kernel of P 1 × 1；The P is that the output of (n-1)th layer of convolutional layer is led to
Road number；Convolution kernel and the convolution kernel of L 3 × 3 of the second sub- convolutional layer including J individual 1 × 1 in the convolutional layer, P=J+L ×
3；The M, P, J, L are the positive integer more than or equal to 1, and the n is the positive integer more than 1 and less than M.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711123049.6A CN107832835A (en) | 2017-11-14 | 2017-11-14 | The light weight method and device of a kind of convolutional neural networks |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201711123049.6A CN107832835A (en) | 2017-11-14 | 2017-11-14 | The light weight method and device of a kind of convolutional neural networks |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107832835A true CN107832835A (en) | 2018-03-23 |
Family
ID=61655520
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201711123049.6A Pending CN107832835A (en) | 2017-11-14 | 2017-11-14 | The light weight method and device of a kind of convolutional neural networks |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107832835A (en) |
Cited By (17)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108682154A (en) * | 2018-06-19 | 2018-10-19 | 上海理工大学 | Congestion in road detecting system based on the analysis of wagon flow state change deep learning |
| CN108710906A (en) * | 2018-05-11 | 2018-10-26 | 北方民族大学 | Real-time point cloud model sorting technique based on lightweight network LightPointNet |
| CN109214511A (en) * | 2018-08-15 | 2019-01-15 | 算丰科技（北京）有限公司 | Data processing method, data processing equipment and electronic equipment |
| CN109492661A (en) * | 2018-09-27 | 2019-03-19 | 桂林电子科技大学 | Traffic sign recognition method based on WWCNN model and TVPDE algorithm |
| CN109543766A (en) * | 2018-11-28 | 2019-03-29 | 钟祥博谦信息科技有限公司 | Image processing method and electronic equipment, storage medium |
| CN110135465A (en) * | 2019-04-22 | 2019-08-16 | 腾讯科技（深圳）有限公司 | Model parameter representation space size estimation method and device, recommended method |
| CN110175262A (en) * | 2019-05-31 | 2019-08-27 | 武汉斗鱼鱼乐网络科技有限公司 | Deep learning model compression method, storage medium and system based on cluster |
| CN110188863A (en) * | 2019-04-30 | 2019-08-30 | 杭州电子科技大学 | A kind of convolution kernel and its compression algorithm of convolutional neural networks |
| CN110364010A (en) * | 2019-08-22 | 2019-10-22 | 三星电子（中国）研发中心 | A kind of navigation methods and systems of predicting road conditions |
| CN110363279A (en) * | 2018-03-26 | 2019-10-22 | 华为技术有限公司 | Image processing method and device based on convolutional neural networks model |
| CN110490295A (en) * | 2018-05-15 | 2019-11-22 | 华为技术有限公司 | A kind of neural network model, data processing method and processing unit |
| CN110647898A (en) * | 2018-06-27 | 2020-01-03 | 北京市商汤科技开发有限公司 | Image processing method, image processing device, electronic equipment and computer storage medium |
| CN110852348A (en) * | 2019-10-18 | 2020-02-28 | 北京迈格威科技有限公司 | Feature map processing method, image processing method and device |
| CN111783898A (en) * | 2020-07-09 | 2020-10-16 | 腾讯科技（深圳）有限公司 | Training method of image recognition model, image recognition method, device and equipment |
| CN112733585A (en) * | 2019-10-29 | 2021-04-30 | 杭州海康威视数字技术股份有限公司 | Image recognition method |
| CN112990454A (en) * | 2021-02-01 | 2021-06-18 | 国网安徽省电力有限公司检修分公司 | Neural network calculation acceleration method and device based on integrated DPU multi-core isomerism |
| WO2021253941A1 (en) * | 2020-06-18 | 2021-12-23 | 华为技术有限公司 | Neural network model training method, image classification method, text translation method and apparatus, and device |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160019454A1 (en) * | 2014-07-18 | 2016-01-21 | James LaRue | J Patrick's Ladder A Machine Learning Enhancement Tool |
| CN106127297A (en) * | 2016-06-02 | 2016-11-16 | 中国科学院自动化研究所 | The acceleration of degree of depth convolutional neural networks based on resolution of tensor and compression method |
| CN106779051A (en) * | 2016-11-24 | 2017-05-31 | 厦门中控生物识别信息技术有限公司 | A kind of convolutional neural networks model parameter processing method and system |
| CN106886975A (en) * | 2016-11-29 | 2017-06-23 | 华南理工大学 | It is a kind of can real time execution image stylizing method |
| CN107239824A (en) * | 2016-12-05 | 2017-10-10 | 北京深鉴智能科技有限公司 | Apparatus and method for realizing sparse convolution neutral net accelerator |
| CN107239825A (en) * | 2016-08-22 | 2017-10-10 | 北京深鉴智能科技有限公司 | Consider the deep neural network compression method of load balancing |
-
2017
- 2017-11-14 CN CN201711123049.6A patent/CN107832835A/en active Pending
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20160019454A1 (en) * | 2014-07-18 | 2016-01-21 | James LaRue | J Patrick's Ladder A Machine Learning Enhancement Tool |
| CN106127297A (en) * | 2016-06-02 | 2016-11-16 | 中国科学院自动化研究所 | The acceleration of degree of depth convolutional neural networks based on resolution of tensor and compression method |
| CN107239825A (en) * | 2016-08-22 | 2017-10-10 | 北京深鉴智能科技有限公司 | Consider the deep neural network compression method of load balancing |
| CN106779051A (en) * | 2016-11-24 | 2017-05-31 | 厦门中控生物识别信息技术有限公司 | A kind of convolutional neural networks model parameter processing method and system |
| CN106886975A (en) * | 2016-11-29 | 2017-06-23 | 华南理工大学 | It is a kind of can real time execution image stylizing method |
| CN107239824A (en) * | 2016-12-05 | 2017-10-10 | 北京深鉴智能科技有限公司 | Apparatus and method for realizing sparse convolution neutral net accelerator |
Non-Patent Citations (2)
| Title |
|---|
| 付华 等: "《智能仪器设计》", 31 August 2007, 国防工业出版社 * |
| 陈敏: "《认知计算导论》", 30 April 2017, 华中科技大学出版社 * |
Cited By (30)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110363279A (en) * | 2018-03-26 | 2019-10-22 | 华为技术有限公司 | Image processing method and device based on convolutional neural networks model |
| CN110363279B (en) * | 2018-03-26 | 2021-09-21 | 华为技术有限公司 | Image processing method and device based on convolutional neural network model |
| CN108710906A (en) * | 2018-05-11 | 2018-10-26 | 北方民族大学 | Real-time point cloud model sorting technique based on lightweight network LightPointNet |
| CN108710906B (en) * | 2018-05-11 | 2022-02-11 | 北方民族大学 | Real-time point cloud model classification method based on lightweight network LightPointNet |
| CN110490295B (en) * | 2018-05-15 | 2022-04-05 | 华为技术有限公司 | Data processing method and processing device |
| CN110490295A (en) * | 2018-05-15 | 2019-11-22 | 华为技术有限公司 | A kind of neural network model, data processing method and processing unit |
| CN108682154B (en) * | 2018-06-19 | 2021-03-16 | 上海理工大学 | Road congestion detection system based on deep learning analysis of traffic flow state change |
| CN108682154A (en) * | 2018-06-19 | 2018-10-19 | 上海理工大学 | Congestion in road detecting system based on the analysis of wagon flow state change deep learning |
| CN110647898A (en) * | 2018-06-27 | 2020-01-03 | 北京市商汤科技开发有限公司 | Image processing method, image processing device, electronic equipment and computer storage medium |
| CN110647898B (en) * | 2018-06-27 | 2022-11-01 | 北京市商汤科技开发有限公司 | Image processing method, image processing device, electronic equipment and computer storage medium |
| CN109214511B (en) * | 2018-08-15 | 2021-08-27 | 算丰科技（北京）有限公司 | Data processing method, data processing device and electronic equipment |
| CN109214511A (en) * | 2018-08-15 | 2019-01-15 | 算丰科技（北京）有限公司 | Data processing method, data processing equipment and electronic equipment |
| CN109492661A (en) * | 2018-09-27 | 2019-03-19 | 桂林电子科技大学 | Traffic sign recognition method based on WWCNN model and TVPDE algorithm |
| CN109492661B (en) * | 2018-09-27 | 2021-08-13 | 桂林电子科技大学 | Traffic sign identification method based on WWCNN model and TVPDE algorithm |
| CN109543766A (en) * | 2018-11-28 | 2019-03-29 | 钟祥博谦信息科技有限公司 | Image processing method and electronic equipment, storage medium |
| CN110135465A (en) * | 2019-04-22 | 2019-08-16 | 腾讯科技（深圳）有限公司 | Model parameter representation space size estimation method and device, recommended method |
| CN110135465B (en) * | 2019-04-22 | 2022-12-09 | 腾讯科技（深圳）有限公司 | Model parameter representation space size estimation method and device and recommendation method |
| CN110188863B (en) * | 2019-04-30 | 2021-04-09 | 杭州电子科技大学 | Convolution kernel compression method of convolution neural network suitable for resource-limited equipment |
| CN110188863A (en) * | 2019-04-30 | 2019-08-30 | 杭州电子科技大学 | A kind of convolution kernel and its compression algorithm of convolutional neural networks |
| CN110175262A (en) * | 2019-05-31 | 2019-08-27 | 武汉斗鱼鱼乐网络科技有限公司 | Deep learning model compression method, storage medium and system based on cluster |
| CN110364010A (en) * | 2019-08-22 | 2019-10-22 | 三星电子（中国）研发中心 | A kind of navigation methods and systems of predicting road conditions |
| CN110852348B (en) * | 2019-10-18 | 2022-09-30 | 北京迈格威科技有限公司 | Feature map processing method, image processing method and device |
| CN110852348A (en) * | 2019-10-18 | 2020-02-28 | 北京迈格威科技有限公司 | Feature map processing method, image processing method and device |
| CN112733585A (en) * | 2019-10-29 | 2021-04-30 | 杭州海康威视数字技术股份有限公司 | Image recognition method |
| CN112733585B (en) * | 2019-10-29 | 2023-09-05 | 杭州海康威视数字技术股份有限公司 | image recognition method |
| WO2021253941A1 (en) * | 2020-06-18 | 2021-12-23 | 华为技术有限公司 | Neural network model training method, image classification method, text translation method and apparatus, and device |
| CN111783898B (en) * | 2020-07-09 | 2021-09-14 | 腾讯科技（深圳）有限公司 | Training method of image recognition model, image recognition method, device and equipment |
| CN111783898A (en) * | 2020-07-09 | 2020-10-16 | 腾讯科技（深圳）有限公司 | Training method of image recognition model, image recognition method, device and equipment |
| CN112990454A (en) * | 2021-02-01 | 2021-06-18 | 国网安徽省电力有限公司检修分公司 | Neural network calculation acceleration method and device based on integrated DPU multi-core isomerism |
| CN112990454B (en) * | 2021-02-01 | 2024-04-16 | 国网安徽省电力有限公司超高压分公司 | Neural network calculation acceleration method and device based on integrated DPU multi-core heterogeneous |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107832835A (en) | The light weight method and device of a kind of convolutional neural networks | |
| CN106297297B (en) | Traffic jam judging method based on deep learning | |
| CN109948425A (en) | A kind of perception of structure is from paying attention to and online example polymerize matched pedestrian's searching method and device | |
| CN108509978A (en) | The multi-class targets detection method and model of multi-stage characteristics fusion based on CNN | |
| CN105488528B (en) | Neural network image classification method based on improving expert inquiry method | |
| CN110321952A (en) | A kind of training method and relevant device of image classification model | |
| CN108416250A (en) | Demographic method and device | |
| WO2019179403A1 (en) | Fraud transaction detection method based on sequence width depth learning | |
| CN105279554B (en) | The training method and device of deep neural network based on Hash coding layer | |
| CN105574550A (en) | Vehicle identification method and device | |
| KR102328734B1 (en) | Method for automatically evaluating labeling reliability of training images for use in deep learning network to analyze images, and reliability-evaluating device using the same | |
| CN109886066A (en) | Fast target detection method based on the fusion of multiple dimensioned and multilayer feature | |
| CN112464911A (en) | Improved YOLOv 3-tiny-based traffic sign detection and identification method | |
| CN109784283A (en) | Based on the Remote Sensing Target extracting method under scene Recognition task | |
| CN112541532B (en) | Target detection method based on dense connection structure | |
| CN108960015A (en) | A kind of vehicle system automatic identifying method and device based on deep learning | |
| CN107092883A (en) | Object identification method for tracing | |
| CN106682681A (en) | Recognition algorithm automatic improvement method based on relevance feedback | |
| CN110399820B (en) | Visual recognition analysis method for roadside scene of highway | |
| CN113034506B (en) | Remote sensing image semantic segmentation method and device, computer equipment and storage medium | |
| CN111832615A (en) | Sample expansion method and system based on foreground and background feature fusion | |
| CN111832484A (en) | Loop detection method based on convolution perception hash algorithm | |
| CN110210550A (en) | Image fine granularity recognition methods based on integrated study strategy | |
| CN113743417A (en) | Semantic segmentation method and semantic segmentation device | |
| CN113298817A (en) | High-accuracy semantic segmentation method for remote sensing 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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180323 |