CN110288030A - Image-recognizing method, device and equipment based on lightweight network model - Google Patents

Image-recognizing method, device and equipment based on lightweight network model Download PDF

Info

Publication number
CN110288030A
CN110288030A CN201910566189.3A CN201910566189A CN110288030A CN 110288030 A CN110288030 A CN 110288030A CN 201910566189 A CN201910566189 A CN 201910566189A CN 110288030 A CN110288030 A CN 110288030A
Authority
CN
China
Prior art keywords
image
node
network model
characteristic
neural networks
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN201910566189.3A
Other languages
Chinese (zh)
Other versions
CN110288030B (en
Inventor
房斌
李婷
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Chongqing University
Original Assignee
Chongqing University
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Chongqing University filed Critical Chongqing University
Priority to CN201910566189.3A priority Critical patent/CN110288030B/en
Publication of CN110288030A publication Critical patent/CN110288030A/en
Application granted granted Critical
Publication of CN110288030B publication Critical patent/CN110288030B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • 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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a kind of image-recognizing method based on lightweight network model, device and equipment.Described image recognition methods, comprising the following steps: S1 obtains target image to be identified；Target image is input in the lightweight network model trained by S2；S3 classifies to target image using the lightweight network model trained.Wherein, the process of the lightweight network model is obtained the following steps are included: S21, constructs the variant convolutional neural networks without full articulamentum；S22 is updated the weight of convolutional layer by softmax classifier to image classification；S23 extracts the feature of image using the updated variant convolutional neural networks of weight again, and is standardized to feature；Feature after standardization is generated characteristic node and enhancing node according to the building method of broadband network, determines final characteristic node, enhancing node number, construct lightweight network model by S24.

Description

Image-recognizing method, device and equipment based on lightweight network model
Technical field
The present invention relates to mode identification technologies, and in particular to a kind of image recognition side based on lightweight network model Method, device, equipment and readable storage medium storing program for executing.
Background technique
When deep neural network is applied to field of image recognition, since deep neural network is related to a large amount of hyper parameter and answers Miscellaneous structure, this complexity make theoretically analysis depth structure extremely difficult, most of work be related to adjusting parameter or More layers is stacked to obtain better precision, although therefore deep neural network precision high calculating time and training time It is long.Article " Broad Learning System:An Effective and Efficient Incremental Learning The broadband network (BLS) proposed in System Without the Need for Deep Architecture " is to be based on The idea of RVFLNN and design, relative to " depth " structure, " width " structure due to coupling not between layers rather than Often succinct, the training process of BLS reduces the dependence to computer and storage resource.Because connecting without multilayer, BLS is not needed Weight is updated using gradient decline, BLS asks matrix pseudoinverse to find required connection weight by ridge regression, reaches in neural network accuracy When less than requiring, by increasing " width " Lai Tisheng precision of network, network is quickly rebuild using Incremental Learning Algorithm, without weight New training process, so calculating speed is significantly better than deep learning.Although the promotion of BLS speed is fairly obvious, it is applied to figure As Classification and Identification field, nicety of grading is not high enough.
Summary of the invention
It is an object of the invention to overcome the above-mentioned deficiency in the presence of the prior art, provide a kind of based on lightweight network Image-recognizing method, device, equipment and the readable storage medium storing program for executing of model, with realize image classification model size, efficiency, resource, Balance above precision, and then realize the fast and accurate true classification of image, it solves depth network and relies on expensive hardware configuration and meter Calculation and time consumption for training, and the problem that broadband network precision is not high.
In order to achieve the above-mentioned object of the invention, the present invention provides following technical schemes:
A kind of image-recognizing method based on lightweight network model, comprising the following steps:
S1 obtains target image to be identified；
Target image is input in the lightweight network model trained by S2；
S3 classifies to target image using the lightweight network model trained.
Wherein, obtain the process of the lightweight network model the following steps are included:
S21 constructs the variant convolutional neural networks without full articulamentum according to the make of convolutional neural networks, described Variant convolutional neural networks include one and its above network layer, and the network layer includes a convolutional layer and a pond Layer；
S22, the image to be classified marked obtain the characteristic image of image by variant convolutional neural networks, and this feature is defeated Enter softmax classifier, by softmax classifier to image classification, foundation classification results and the true value for having marked image It is updated by weight of the loss function to convolutional layer；
S23 extracts the feature of image using the updated variant convolutional neural networks of weight again, and marks to feature Quasi-ization processing；
Feature after standardization is generated characteristic node and enhancing node according to the building method of broadband network, determined by S24 Final characteristic node, enhancing node number, construct lightweight network model.
Preferably, convolutional layer convolution operation process is as follows in the step S21:
Assuming thatFor raw image data collection, Xi is i-th of image, i=1,2 ... n；L indicates the network number of plies, And the size of l layers of Convolution Filter is kl*kl, depth dl, the moving step length of Convolution Filter is sl；
I-th of image XiA series of convolution operations carried out indicate are as follows:
Wherein, l and l-1 indicates the convolutional layer number of plies, and l is current layer, and l-1 is preceding layer；Indicate l layers of convolutional layer output Image,Indicate the image that i-th of image exports after the processing of pond layer in l-1 layer network layer, andFor original Beginning image Xi；W is the weight of convolutional layer, and b is the biasing of convolutional layer, and W and b is randomly generated, i.e. W(l)For l layers of convolutional layer Weight matrix, b(l)For the bias matrix of l layers of convolutional layer；For convolution operation.
Preferably, the pond pond Hua Ceng operating process is as follows in the step S21:
Using the pond mode of operation being maximized, the step-length of each pond layer is tl, therefore, the spy after a convolution operation A channel poolization for levying picture operates unfolding calculation are as follows:
Wherein, C is the characteristic image exported after pondization operates.
Preferably, the detailed process of the step S22 is as follows:
The character image data of training sample by multilayer convolution, pondization operation output is expressed asIts In, C ' is the characteristic pattern image set exported after multilayer convolution, pondization operation, and n is image number, vector U 'iFor i-th of image By the characteristic image that variant CNN is extracted, and i ∈ [1, n]；Image K class to be divided at this time will pass through multilayer convolution, Chi Huacao The characteristic image of the output of work is connected to K node entirely, indicates are as follows:
Yy=WYC′+bY, y ∈ [1, K]；
Wherein YyFor the output of y-th of node, WYFor the weight of node operation, bYFor the biasing of node operation；
After being connected to K node entirely, classified with softmax classifier to operation result, softmax algorithm are as follows:
Wherein, SyIndicate that image is divided into the probability of y-th of classification, ayThe value classified for y-th, akClassify for k-th Value, k ∈ [1, K]；ayThe output Y of y-th obtained of node is as connected entirelyy。
Using cross entropy L as loss function, cross entropy L is indicated are as follows:
Based on cross entropy loss function, weight update is carried out to variant convolutional neural networks by Adam algorithm.
Preferably, the detailed process of the step S23 is as follows:
The feature of image is extracted again using the updated variant convolutional neural networks of weight, and feature is standardized Processing, standardized detailed process are as follows:
Wherein U 'iRefer to i-th of the image extracted by the updated variant convolutional neural networks trained of weight Characteristic image, μiFor the mean value of the characteristic image of i-th of image of output, σ is the characteristic image of i-th of image of output Variance；U″iFor the characteristic image of i-th of image after standardization；
Standardized feature set representations after the characteristic image standardization of all images are as follows:
Preferably, the detailed process of the step S24 is as follows:
Broadband network characteristic node indicates are as follows:
Wherein, ZrIndicate the characteristic node of broadband network；φ is arbitrary function；C " is the characteristic image after standardization； For the random weight coefficient of the characteristic node with appropriate dimension；It is characterized the biasing of node, e is the feature of broadband network The number of node；
Define ZE=[Z1, Z2..., Ze]；
Then enhance node to be expressed as:
Define HF=[H1, H2..., Hf]；
Wherein, HjIndicate the characteristic node of broadband network；ζ is activation primitive；For the enhancing section with appropriate dimension The random weight coefficient of point；For the biasing for enhancing node, f is the number of the characteristic node of broadband network；
It is a kind of to be generally designated as based on lightweight network structure:
A kind of pattern recognition device based on lightweight network model, including target image acquisition module, target image are defeated Enter module, Classification and Identification module and object module and obtains module；
Wherein, target image obtains module, for obtaining target image to be identified；
Target image input module, for target image to be input in target lightweight network model；The lightweight Network model is obtained by the variant convolutional neural networks and broadband network construction for removing full articulamentum；
Classification and Identification module is known for carrying out classification processing to target image using target lightweight network model Other result；
Object module obtains module, comprising:
Model construction unit, for constructing the variant convolutional neural networks without full articulamentum, and the structure based on broadband network It builds step and obtains broadband network characteristic node and enhancing node, construct lightweight network model；
Loss function is inserted into unit, for the insertion loss function in variant convolutional neural networks；
Training unit instructs variant convolutional neural networks for utilizing loss function combination softmax classifier Practice, undated parameter, obtains object module.
A kind of image recognition apparatus based on lightweight network model, comprising:
Memory, for storing computer program；
Processor realizes the above-mentioned image recognition side based on lightweight network model when for executing the computer program The step of method.
A kind of readable storage medium storing program for executing, which is characterized in that computer program, the meter are stored on the readable storage medium storing program for executing Calculation machine program realizes the step of above-mentioned image-recognizing method based on lightweight network model when being executed by processor.
Compared with prior art, characteristics of image, input beneficial effects of the present invention: are extracted by variant convolutional neural networks The network structure of construction step invention based on broadband network, obtains lightweight network model, which can The dependence to image recognition processes to computer storage resource is reduced, the training time can be effectively reduced, and improves image point Class precision realizes image classification model size, efficiency, resource, the balance above precision.
Detailed description of the invention:
Fig. 1 is the lightweight network architecture figure of exemplary embodiment of the present 1；
Fig. 2 is the flow chart of the image-recognizing method based on lightweight network model of exemplary embodiment of the present 1；
Fig. 3 is the step S2's of the image-recognizing method based on lightweight network model of exemplary embodiment of the present 1 Detail flowchart；
Fig. 4 is the MNIST that the image-recognizing method in exemplary embodiment of the present 2 based on lightweight network model extracts The characteristic effect figure of data set；
Fig. 5 is the structural representation of the pattern recognition device based on lightweight network model in exemplary embodiment of the present 3 Figure；
Fig. 6 is the structural representation of the image recognition apparatus based on lightweight network model in exemplary embodiment of the present 4 Figure；
Fig. 7 is the specific structure of the image recognition apparatus based on lightweight network model in exemplary embodiment of the present 4 Schematic diagram.
Specific embodiment
Below with reference to test example and specific embodiment, the present invention is described in further detail.But this should not be understood It is all that this is belonged to based on the technology that the content of present invention is realized for the scope of the above subject matter of the present invention is limited to the following embodiments The range of invention.
Embodiment 1
As shown in Figure 1 to Figure 3, the present embodiment provides a kind of image-recognizing methods based on lightweight network model, specifically The following steps are included:
S1 obtains target image to be identified；
Target image is input in the lightweight network model trained by S2；
S3 classifies to target image using the lightweight network model trained.
Wherein, obtain the process of the lightweight network model the following steps are included:
S21 constructs the variant convolutional neural networks (CNN) without full articulamentum according to the make of convolutional neural networks, The variant convolutional neural networks include one and its above network layer, and the network layer includes a convolutional layer and a pond Change layer；
S22 inputs the image to be classified marked；The image to be classified marked is obtained by variant convolutional neural networks The feature of image, this feature input softmax classifier, by softmax classifier to image classification, according to classification results and The true value for having marked image is updated by weight of the loss function to convolutional layer；
S23 extracts the feature of image using the updated variant convolutional neural networks of weight again, and marks to feature Quasi-ization processing；
Feature after standardization is generated a certain number of characteristic nodes according to the building method of broadband network by S24 With enhancing node, determines final characteristic node, enhancing node number, construct lightweight network model.
The convolutional neural networks for removing full articulamentum are used for image characteristics extraction by above-mentioned steps, pass through the power of convolution operation Value is shared and pondization operates the order of magnitude for reducing network parameter, enhances the feature of broadband network BLS input data.Pass through combination Variant convolutional neural networks and broadband network construct a kind of light-weighted network structure, realize fast and accurate true point of image Class.
LeNet, AlexNet, VggNet, ResNet etc. are the common network structure model in convolutional neural networks field, it Make have certain difference, and be adapted to data set of different sizes.The present embodiment is according to image to be classified data set Size selects suitable convolutional layer and pond layer number, the size of convolutional layer filter, convolutional layer filter moving step length, Chi Hua The mode of operation, to construct the variant convolutional neural networks without full articulamentum.Wherein, pondization operation is maximized and is averaged Two kinds of modes of operation of value.
Convolutional layer convolution operation process is as follows in step S21:
Assuming thatFor raw image data collection, XiFor i-th of image, i=1,2 ... n；L indicates the network number of plies, And the size of l layers of Convolution Filter is kl*kl, depth dl, the moving step length of Convolution Filter is sl；
I-th of image XiA series of convolution operations carried out indicate are as follows:
Wherein, l and l-1 indicates the convolutional layer number of plies, and l is current layer, and l-1 is preceding layer；Indicate the volume of l layer network layer The image of lamination output,Indicate the image that i-th of image exports after the processing of pond layer in l-1 layer network layer, AndFor original image Xi；W is the weight of convolutional layer, and b is the biasing of convolutional layer, and W and b is randomly generated, i.e. W(l)It is l layers The weight matrix of convolutional layer, b(l)For the bias matrix of l layers of convolutional layer；For convolution operation.
One channel convolution operation unfolding calculation of one image are as follows:
Wherein, WpqFor the weight coefficient in weight matrix, wherein (1, c) p ∈, q ∈ (1, c), the size of convolution kernel at this time For c*c；XghFor input convolutional layer image slices vegetarian refreshments value, wherein (1, m) g ∈, h ∈ (1, m), image dimension m*m；B is The biasing of convolutional layer；S is the moving step length of Convolution Filter.Since before image recognition resolution sizes would generally be carried out to image The processing such as change, enhancing and denoising, the image for inputting network is generally square, therefore image shape is square herein, Resolution sizes are m*m, but the image of the application can be other shapes, and those skilled in the technology concerned are not departing from this In the case where the principle and range of invention, various replacements, modification and the improvement made should be included in protection model of the invention Within enclosing.
It is m/s that the height new_height and width new_width of image characteristic matrix are exported after convolution operation, in order to Matrix convenience of calculation, needs to expand the matrix size of the image of input convolution operation, and pixel numerical value to be supplemented is denoted as 0；It is defeated Entering image array in height needs the pixel quantity expanded are as follows:
Pad_needed_height=(new_height-1) * s+c-m；
So, the pixel quantity pad_top expanded and lower section is needed to need the pixel quantity expanded above input matrix Pad_bottom calculation is as follows:
Pad_top=pad_neededd_height/2, pad_bottom=pad_needed_height-pad_top；
The pixel quantity pad_left and the right that the input matrix left side need to be expanded need the pixel quantity pad_ expanded Right calculation is as follows:
Pad_left=pad_top, pad_right=pad_bottom.
Hua Ceng pond operating process in pond is as follows in step S21:
The operation of the present embodiment pondization is t using the mode of operation being maximized, the step-length of l layer network layer pond layerl, and the The moving step length of l layer network Convolution Filter is sl, therefore, a channel poolization of the feature image after a convolution operation is grasped Make unfolding calculation are as follows:
Wherein, C is the characteristic image exported after pondization operates.
The detailed process of step S22 is as follows.
The character image data of training sample by multilayer convolution, pondization operation output is expressed asIts In, C ' is the characteristic pattern image set exported after multilayer convolution, pondization operation, and n is image number, vector U 'iFor i-th of image The pixel value collection of each pixel of characteristic image by variant CNN extraction, and i ∈ [1, n].Image K class to be divided at this time, will be through Cross multilayer convolution, the characteristic image for the output that pondization operates is connected to K node entirely, expression are as follows:
Yy=WYC′+bY, y ∈ [1, K],
Wherein YyFor the output of y-th of node, WYFor the weight of node operation, bYFor the biasing of node operation.
After being connected to K node entirely, classified with softmax classifier to operation result, softmax algorithm are as follows:
Wherein, SyIndicate that image is divided into the probability of y-th of classification, ayThe value classified for y-th, akClassify for k-th Value, k ∈ [1, K]；ayThe output Y of y-th obtained of node is as connected entirelyy。
Using cross entropy L as loss function, cross entropy L is indicated the loss function of this implementation are as follows:
Operation values based on loss function carry out weight update to variant convolutional neural networks by using Adam algorithm. During constructing variant convolutional neural networks, weight and offset parameter are randomly generated, therefore not can guarantee convolution Chi Huacao It extracts and is characterized in preferably after work, in order to solve this problem, network weight update is carried out using Adam algorithm.Adam is calculated The single order moments estimation and second order moments estimation that method passes through stochastic gradient are the independent adaptivity learning rate of different parameter designings, It is advantageous in non-convex optimization problem, Adam algorithm and other existing optimization algorithms (such as Gradient Descent, Adadelta and Adagrad algorithm) it compares, effect of optimization is more preferable, is classified using the network image after Adam algorithm optimization smart Degree is high.
The Adam algorithm of more neomorph convolutional neural networks weight are as follows:
zλ-1=β1zλ-2+(1-β1)f′(θλ-1),
vλ-1=β2vλ-2+(1-β2)f′(θλ-1)2；
Wherein λ indicates the number of iterations；α is hyper parameter learning rate；β1、β2For the exponential decay rate of hyper parameter moments estimation, use In the attenuation rate for controlling mobile mean value；ε is smooth item；θ is aleatory variable；Z, v is the null vector being initialised, for variable θ Not to zero offset, deviation correction is done to variable θ, passes through the z after calculating deviation correctionλ、vλTo offset these deviations；θλIndicate λ The θ vector of secondary iteration.
In the present embodiment, parameter beta1、β2Value is 0.9,0.999 respectively, and smooth item ε value is 10-8, learning rate α is in net It is finely adjusted when network training.Small lot sample training is chosen when weight updates, as far as possible reduction network training the number of iterations.
It is minimized using gradient descent algorithm undated parameter by gradient descent method come iterative solution step by step Loss function and corresponding model parameter value, specific calculating process is as follows:
Derivative of the loss function about convolutional layer weight W are as follows:
Derivative of the loss function about convolutional layer biasing b are as follows:
WhereinCalculation formula is as follows:
For in l layers of weight matrix pth row q column weight coefficient,The image of convolutional layer is inputted for l layers G row h column pixel value, b(l)For l layers of biasing,For the g row h of l layers of convolutional layer treated image The value of the pixel of column.
According to the make of convolutional neural networks, the variant convolutional neural networks without full articulamentum, the variant are constructed Convolutional neural networks include convolutional layer and pond layer, are shared by the weight of convolution operation and pondization operation reduces network parameter The order of magnitude, preferably extraction characteristics of image；Then by softmax classifier to image classification, by using Adam algorithm base Weight update is carried out to variant convolutional neural networks by gradient descent method in the operation values of cross entropy loss function.Weight updates Variant convolutional neural networks feature extraction afterwards is more accurate.
Extract the feature of image in step S23 again using the updated variant convolutional neural networks of weight, and to feature It is standardized, standardized detailed process is as follows:
Wherein U 'iRefer to the characteristic image by the updated variant CNN trained of weight i-th of the image extracted The pixel value collection of each pixel, μiFor the mean value of the characteristic image of i-th of image of output, σ is the spy of i-th of image of output Levy the variance of image；U″iFor the pixel value of each pixel of characteristic image of i-th of image after standardization.
Standardized feature set representations after the characteristic image standardization of all images are as follows:
Step S24 by the feature after standardization according to the building method of broadband network generate a certain number of characteristic nodes and Enhance node, determines final characteristic node, enhancing node number with grid data service, construct lightweight network model.Width The characteristic node of network is used to extract the feature of image；And enhance node and increase the non-linear of whole network, for classifying.In detail Thin construction process is as follows:
Broadband network characteristic node indicates are as follows:
Wherein, ZrIndicate the characteristic node of broadband network；φ is arbitrary function；C " is the characteristic image after standardization； For the random weight coefficient of the characteristic node with appropriate dimension；It is characterized the biasing of node, e is the feature of broadband network The number of node.
Define ZE=[Z1, Z2..., Ze]；
Then enhance node to be expressed as:
Define HF=[H1, H2..., Hf]
Wherein, HjIndicate the characteristic node of broadband network；ζ is activation primitive；For the enhancing section with appropriate dimension The random weight coefficient of point；For the biasing for enhancing node.
It is a kind of to be generally designated as based on lightweight network structure:
By above-mentioned steps, characteristics of image is extracted by variant convolutional neural networks (CNN), feature input is based on width The lightweight network model of the construction step invention of network, is rapidly performed by image recognition.Based on light weight described in the present embodiment The image-recognizing method for changing network model can reduce the dependence to computer storage resource, can effectively reduce the training time, And image classification accuracy is improved, efficiency, resource and precision three requirements are better balanced.
Embodiment 2
Image-recognizing method described in embodiment 1 based on lightweight network model can be widely used in image recognition neck Domain, the present embodiment test using MNIST data set and training network, detailed process are as follows:
Variant convolutional neural networks (CNN) is constructed to extract characteristics of image.It is with reference to building variant with LeNet network structure CNN, the first convolutional layer, the first pond layer, the second convolutional layer and the second pond layer connect in order in variant convolutional neural networks It connects, the size of the Convolution Filter of the first convolutional layer and the second convolutional layer is all 5*5, and moving step length is all 1, wherein the first convolution The depth of the filter of layer is 32, and the depth of the filter of the second convolutional layer is 64；First pond layer and the second pond layer are all adopted With maximum Chi Huafa, and the moving step length of pond layer filter is all 2, and the last layer pond result is connected to certain amount Node on, construct the variant CNN without full articulamentum, then classified with softmax classifier, cross entropy as loss letter Number.
The image that 60000 sizes are 28*28*1 in MNIST data set is inputted as the training sample of variant CNN, note ForIt is as follows according to convolution pond operation processing image, concrete processing procedure described in embodiment 1:
XiBy the calculating process expansion in one channel of filter of the first convolutional layer are as follows:
The result in above-mentioned channel becomes by the first pond layer filter:
Therefore, on the data set arbitrary image last output result are as follows:
The last layer pond result of all training sample images:
It is connected on 10 nodes entirely, corresponding output
The CNN variant without full articulamentum is constructed, is then classified with softmax classifier, cross entropy is as loss letter Number, carries out network weight update with Adam algorithm.
Weight updates chooses 100 samples every time, and iteration 1000 times, the number of iterations here can be according to promotion precision It needs and considers that time cost suitably increases or decreases.After iteration 1000 times, it is assumed that network weight is W ' at this time, is biased to b ', Extract the feature of image, the feature set representations of all images of extraction again using the updated variant convolutional neural networks of weight Are as follows:
The feature that variant CNN is extracted than simply projecting, direction, center of gravity will be more scientific.It can use different The capability of fitting of the size control overall model of convolution, pond and the feature vector finally exported.It can be reduced in over-fitting The dimension of feature vector the output dimension of convolutional layer can be improved in poor fitting, more compared to other feature extracting methods Flexibly.
The set of image characteristics extracted to the updated variant convolutional neural networks of weight is standardized.
Image standardization is by data by going mean value to realize the processing of centralization, according to convex optimum theory and data probability It is distributed relevant knowledge, data center meets data distribution rule, it is easier to obtain extensive effect after training.Each image Characteristics of imageStandardized feature collection after the characteristic image standardization of all images are as follows:
The characteristics of image that the variant CNN trained is extracted constructs lightweight network structure as input.
Feature after standardization generates a certain number of characteristic nodes and enhancing according to the building method of broadband network (BLS) Node determines that the number of characteristic node is 10 with grid data service, and the number for enhancing node is 11000, and characteristics of image is raw At characteristic node are as follows:
Define ZE=[Z1, Z2..., Z10]；
Then enhance node to be expressed as:
It is a kind of to be generally designated as based on lightweight network structure:
Retain training pattern, classifies to MNIST test set image.
Fig. 4 show the characteristic effect figure that variant CNN is extracted in MNIST data set, and specially handwritten numeral 7 is passed through respectively The characteristic effect figure that first convolutional layer, the first pond layer, the second convolutional layer and the second pond layer obtain.
Efficiency, resource and essence is better balanced in image-recognizing method provided by the embodiment based on lightweight network model Three requirements are spent, Detailed Experimental parameter is as shown in Table 1, respectively shows deep neural network (LeNet5), width nerve net The experiment effect of network (BLS) and lightweight network application when MNIST data set carries out image recognition.Experiment effect includes essence Degree, training time and testing time.
Table 1
Network model classification Precision (%) Training time (s) Testing time (s)
Deep neural network (LeNet5) 98.96 598.21 4.92
Width neural network (BLS) 98.85 142.91 3.67
Lightweight network 99.27 359.22 4.11
By the experimental data of table 1 it is found that the image-recognizing method provided by the embodiment based on lightweight network model more Balance efficiency, resource and precision three requirements, image recognition effect are more preferable well.
Embodiment 3
Corresponding to above method embodiment, the present embodiment additionally provides a kind of image knowledge based on lightweight network model Other device, the pattern recognition device described below based on lightweight network model are based on lightweight network mould with above-described The image-recognizing method of type can correspond to each other reference.
Shown in Figure 5, which comprises the following modules: target image obtains module 101, target image input module 102, Classification and Identification module 103 and object module obtain module 104；
Wherein, target image obtains module 101, for obtaining target image to be identified；
Target image input module 102, for target image to be input in target lightweight network model；The light weight Change network model to obtain by the variant convolutional neural networks and broadband network construction for removing full articulamentum；
Classification and Identification module 103 is obtained for carrying out classification processing to target image using target lightweight network model Recognition result；
Object module obtains module 104, comprising:
Model construction unit, for constructing the variant convolutional neural networks without full articulamentum, and the structure based on broadband network It builds step and obtains broadband network characteristic node and enhancing node, construct lightweight network model；
Loss function is inserted into unit, for the insertion loss function in variant convolutional neural networks；
Training unit instructs variant convolutional neural networks for utilizing loss function combination softmax classifier Practice, undated parameter, obtains object module.
Using device provided by the embodiment of the present invention, target image to be identified is obtained；Target image is input to point Class identification module obtains recognition result.
Target image to be identified is obtained, target image is input in lightweight network model.The lightweight network mould Type is obtained based on variant convolutional neural networks and broadband network without full articulamentum.That is, the lightweight network model is logical The weight for crossing convolution operation is shared and the order of magnitude of pondization operation reduction network parameter, enhances broadband network BLS input data Feature.Then, classification processing is carried out to target image using Classification and Identification module, the identification knot of target image can be obtained Fruit.Since lightweight network model used by Classification and Identification module is based on the variant convolutional neural networks without full articulamentum It is obtained with broadband network, the dependence to computer storage resource can be reduced, the training time can be effectively reduced, and improve figure As nicety of grading, efficiency, resource and precision three requirements are better balanced.
In a kind of specific embodiment of the invention, loss function is inserted into unit, is specifically used in variant convolutional Neural Cross entropy loss function is inserted into network.
Embodiment 4
Corresponding to above method embodiment, the present embodiment additionally provides a kind of image knowledge based on lightweight network model Other equipment, a kind of image recognition apparatus based on lightweight network model described below and above-described one kind are based on light weight The image-recognizing method for changing network model can correspond to each other reference.
Shown in Figure 6, being somebody's turn to do the image recognition apparatus based on lightweight network model includes:
Memory D1, for storing computer program；
Processor D2, realized when for executing computer program above method embodiment based on lightweight network model The step of image-recognizing method.
Specifically, referring to FIG. 7, being the tool of the image recognition apparatus provided in this embodiment based on lightweight network model Body structural schematic diagram, the image recognition apparatus based on lightweight network model can generate bigger because of configuration or performance difference Difference, may include one or more processors (central processing units, CPU) 322 (for example, one A or more than one processor) and memory 332, storage Jie of one or more storage application programs 342 or data 344 Matter 330 (such as one or more mass memory units).Wherein, memory 332 and storage medium 330 can be of short duration deposit Storage or persistent storage.The program for being stored in storage medium 330 may include one or more modules (diagram does not mark), often A module may include to the series of instructions operation in data processing equipment.Further, central processing unit 322 can be set It is set to and is communicated with storage medium 330, storage medium 330 is executed on the image recognition apparatus 301 based on lightweight network model In series of instructions operation.
Image recognition apparatus 301 based on lightweight network model can also include one or more power supplys 326, one A or more than one wired or wireless network interface 350, one or more input/output interfaces 358, and/or, one or More than one operating system 341.For example, Windows ServerTM, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM Deng.
Step in image-recognizing method as described above based on lightweight network model can be by being based on lightweight The structure of the image recognition apparatus of network model is realized.
Embodiment 5
Corresponding to above method embodiment, the present embodiment additionally provides a kind of readable storage medium storing program for executing, and described below one Kind of readable storage medium storing program for executing can correspond to each other reference with a kind of above-described image-recognizing method based on lightweight network model.
A kind of readable storage medium storing program for executing is stored with computer program on readable storage medium storing program for executing, and computer program is held by processor The step of image-recognizing method based on lightweight network model of above method embodiment is realized when row.
The readable storage medium storing program for executing be specifically as follows USB flash disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), the various program storage generations such as random access memory (Random Access Memory, RAM), magnetic or disk The readable storage medium storing program for executing of code.
The above, the only detailed description of the specific embodiment of the invention, rather than limitation of the present invention.The relevant technologies The technical staff in field is not in the case where departing from principle and range of the invention, various replacements, modification and the improvement made It should all be included in the protection scope of the present invention.

Claims (9)

1. a kind of image-recognizing method based on lightweight network model, which comprises the following steps:
S1 obtains target image to be identified；
Target image is input in the lightweight network model trained by S2；
S3 classifies to target image using the lightweight network model trained；
Wherein, obtain the process of the lightweight network model the following steps are included:
S21 constructs the variant convolutional neural networks without full articulamentum, the variant according to the make of convolutional neural networks Convolutional neural networks include one and its above network layer, and the network layer includes a convolutional layer and a pond layer；
S22, the image to be classified marked obtain the characteristic image of image, this feature input by variant convolutional neural networks Softmax classifier is logical according to classification results and the true value for having marked image by softmax classifier to image classification Loss function is crossed to be updated the weight of convolutional layer；
S23 extracts the feature of image using the updated variant convolutional neural networks of weight again, and is standardized to feature Processing；
Feature after standardization is generated characteristic node and enhancing node according to the building method of broadband network, determined final by S24 Characteristic node, enhancing node number, construct lightweight network model.
2. the image-recognizing method according to claim 1 based on lightweight network model, which is characterized in that the step Convolutional layer convolution operation process is as follows in S21:
Assuming thatFor raw image data collection, XiFor i-th of image, i=1,2 ... n.；The l expression network number of plies, and the The size of l layers of Convolution Filter is kl*kl, depth dl, the moving step length of Convolution Filter is sl；
I-th of image XiA series of convolution operations carried out indicate are as follows:
Wherein, l and l-1 indicates the convolutional layer number of plies, and l is current layer, and l-1 is preceding layer；Indicate the figure of l layers of convolutional layer output Picture,Indicate the image that i-th of image exports after the processing of pond layer in l-1 layer network layer, andFor original graph As Xi；W is the weight of convolutional layer, and b is the biasing of convolutional layer, and W and b is randomly generated, i.e. W(l)For the weight of l layers of convolutional layer Matrix, b(l)For the bias matrix of l layers of convolutional layer；For convolution operation.
3. the image-recognizing method according to claim 1 based on lightweight network model, which is characterized in that the step Hua Ceng pond operating process in pond is as follows in S21:
Using the pond mode of operation being maximized, the step-length of each pond layer is tl, therefore, the characteristic pattern after a convolution operation One channel poolization of piece operates unfolding calculation are as follows:
Wherein, C is the characteristic image exported after pondization operates.
4. the image-recognizing method according to claim 1 based on lightweight network model, which is characterized in that the step The detailed process of S22 is as follows:
The character image data of training sample by multilayer convolution, pondization operation output is expressed asWherein, C ' For the characteristic pattern image set exported after multilayer convolution, pondization operation, n is image number, vector U 'iPass through for i-th of image The characteristic image that variant CNN is extracted, and i ∈ [1, n]；Image K class to be divided at this time will be operated by multilayer convolution, pondization The characteristic image of output is connected to K node entirely, indicates are as follows:
Yy=WYC′+bY, y ∈ [1, K]；
Wherein YyFor the output of y-th of node, WYFor the weight of node operation, bYFor the biasing of node operation；
After being connected to K node entirely, classified with softmax classifier to operation result, softmax algorithm are as follows:
Wherein, SyIndicate that image is divided into the probability of y-th of classification, ayThe value classified for y-th, akThe value classified for k-th, k ∈ [1, K]；ayThe output Y of y-th obtained of node is as connected entirelyy；
Using cross entropy L as loss function, cross entropy L is indicated are as follows:
Based on cross entropy loss function, weight update is carried out to variant convolutional neural networks by Adam algorithm.
5. the image-recognizing method according to claim 1 based on lightweight network model, which is characterized in that the step The detailed process of S23 is as follows:
It extracts the feature of image again using the updated variant convolutional neural networks of weight, and place is standardized to feature Reason, standardized detailed process are as follows:
Wherein U 'iRefer to the feature of i-th of the image extracted by the updated variant convolutional neural networks trained of weight Image, μiFor the mean value of the characteristic image of i-th of image of output, σ is the variance of the characteristic image of i-th of image of output； U″iFor the characteristic image of i-th of image after standardization；
Standardized feature set representations after the characteristic image standardization of all images are as follows:
6. the image-recognizing method according to claim 1 based on lightweight network model, which is characterized in that the step The detailed process of S24 is as follows:
Broadband network characteristic node indicates are as follows:
Wherein, ZrIndicate the characteristic node of broadband network；φ is arbitrary function；C " is the characteristic image after standardization；For tool There is the random weight coefficient of the characteristic node of appropriate dimension；It is characterized the biasing of node, e is the characteristic node of broadband network Number；
Define ZE=[Z1, Z2..., Ze]；
Then enhance node to be expressed as:
Define HF=[H1, H2..., Hf]；
Wherein, HjIndicate the characteristic node of broadband network；ζ is activation primitive；For with appropriate dimension enhancing node with Machine weight coefficient；For the biasing for enhancing node, f is the number of the characteristic node of broadband network；
It is a kind of to be generally designated as based on lightweight network structure:
7. a kind of pattern recognition device based on lightweight network model, which is characterized in that obtain module, mesh including target image Logo image input module, Classification and Identification module and object module obtain module；
Wherein, target image obtains module, for obtaining target image to be identified；
Target image input module, for target image to be input in target lightweight network model；The lightweight network Model is obtained by the variant convolutional neural networks and broadband network construction for removing full articulamentum；
Classification and Identification module obtains identification knot for carrying out classification processing to target image using target lightweight network model Fruit；
Object module obtains module, comprising:
Model construction unit, the building step for constructing the variant convolutional neural networks without full articulamentum, and based on broadband network It is rapid to obtain broadband network characteristic node and enhancing node, construct lightweight network model；
Loss function is inserted into unit, for the insertion loss function in variant convolutional neural networks；
Training unit is trained variant convolutional neural networks, more for utilizing loss function combination softmax classifier New parameter obtains object module.
8. a kind of image recognition apparatus based on lightweight network model characterized by comprising
Memory, for storing computer program；
Processor is realized when for executing the computer program and is based on lightweight network as described in claim l to 6 any one The step of image-recognizing method of model.
9. a kind of readable storage medium storing program for executing, which is characterized in that be stored with computer program, the calculating on the readable storage medium storing program for executing The image recognition based on lightweight network model as described in any one of claim 1 to 6 is realized when machine program is executed by processor The step of method.
CN201910566189.3A 2019-06-27 2019-06-27 Image identification method, device and equipment based on lightweight network model Active CN110288030B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201910566189.3A CN110288030B (en) 2019-06-27 2019-06-27 Image identification method, device and equipment based on lightweight network model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201910566189.3A CN110288030B (en) 2019-06-27 2019-06-27 Image identification method, device and equipment based on lightweight network model

Publications (2)

Publication Number Publication Date
CN110288030A true CN110288030A (en) 2019-09-27
CN110288030B CN110288030B (en) 2023-04-07

Family

ID=68007719

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201910566189.3A Active CN110288030B (en) 2019-06-27 2019-06-27 Image identification method, device and equipment based on lightweight network model

Country Status (1)

Country Link
CN (1) CN110288030B (en)

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110837890A (en) * 2019-10-22 2020-02-25 西安交通大学 Weight value fixed-point quantization method for lightweight convolutional neural network
CN110909796A (en) * 2019-11-22 2020-03-24 浪潮电子信息产业股份有限公司 Image classification method and related device
CN110929652A (en) * 2019-11-26 2020-03-27 天津大学 Handwritten Chinese character recognition method based on LeNet-5 network model
CN110956202A (en) * 2019-11-13 2020-04-03 重庆大学 Image training method, system, medium and intelligent device based on distributed learning
CN111476138A (en) * 2020-03-31 2020-07-31 万翼科技有限公司 Construction method and identification method of building drawing component identification model and related equipment
CN111614358A (en) * 2020-04-30 2020-09-01 北京的卢深视科技有限公司 Method, system, device and storage medium for feature extraction based on sub-channel quantization
CN112070100A (en) * 2020-09-11 2020-12-11 深圳力维智联技术有限公司 Image feature recognition method and device based on deep learning model and storage medium
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
CN112861896A (en) * 2019-11-27 2021-05-28 北京沃东天骏信息技术有限公司 Image identification method and device
CN112906829A (en) * 2021-04-13 2021-06-04 成都四方伟业软件股份有限公司 Digital recognition model construction method and device based on Mnist data set
CN113205177A (en) * 2021-04-25 2021-08-03 广西大学 Electric power terminal identification method based on incremental collaborative attention mobile convolution
CN113642592A (en) * 2020-04-27 2021-11-12 武汉Tcl集团工业研究院有限公司 Training method of training model, scene recognition method and computer equipment
CN114363477A (en) * 2021-12-30 2022-04-15 上海网达软件股份有限公司 Method and system for video self-adaptive sharpening based on sliding window weight regression
CN114444622A (en) * 2022-04-11 2022-05-06 中国科学院微电子研究所 Fruit detection system and method based on neural network model
CN114781650A (en) * 2022-04-28 2022-07-22 北京百度网讯科技有限公司 Data processing method, device, equipment and storage medium

Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106529578A (en) * 2016-10-20 2017-03-22 中山大学 Vehicle brand model fine identification method and system based on depth learning
CN107657233A (en) * 2017-09-28 2018-02-02 东华大学 Static sign language real-time identification method based on modified single multi-target detection device
CN108304821A (en) * 2018-02-14 2018-07-20 广东欧珀移动通信有限公司 Image-recognizing method and device, image acquiring method and equipment, computer equipment and non-volatile computer readable storage medium storing program for executing
CN108470320A (en) * 2018-02-24 2018-08-31 中山大学 A kind of image stylizing method and system based on CNN
CN108564555A (en) * 2018-05-11 2018-09-21 中北大学 A kind of digital image noise reduction method based on NSST and CNN
CN108717680A (en) * 2018-03-22 2018-10-30 北京交通大学 Spatial domain picture steganalysis method based on complete dense connection network
CN109086806A (en) * 2018-07-16 2018-12-25 福州大学 A kind of IOT portable device visual identity accelerated method based on low resolution, compressed image
CN109492766A (en) * 2018-11-07 2019-03-19 西安交通大学 A kind of width learning method based on minimum P norm
US20190124045A1 (en) * 2017-10-24 2019-04-25 Nec Laboratories America, Inc. Density estimation network for unsupervised anomaly detection
CN109886971A (en) * 2019-01-24 2019-06-14 西安交通大学 A kind of image partition method and system based on convolutional neural networks

Patent Citations (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106529578A (en) * 2016-10-20 2017-03-22 中山大学 Vehicle brand model fine identification method and system based on depth learning
CN107657233A (en) * 2017-09-28 2018-02-02 东华大学 Static sign language real-time identification method based on modified single multi-target detection device
US20190124045A1 (en) * 2017-10-24 2019-04-25 Nec Laboratories America, Inc. Density estimation network for unsupervised anomaly detection
CN108304821A (en) * 2018-02-14 2018-07-20 广东欧珀移动通信有限公司 Image-recognizing method and device, image acquiring method and equipment, computer equipment and non-volatile computer readable storage medium storing program for executing
CN108470320A (en) * 2018-02-24 2018-08-31 中山大学 A kind of image stylizing method and system based on CNN
CN108717680A (en) * 2018-03-22 2018-10-30 北京交通大学 Spatial domain picture steganalysis method based on complete dense connection network
CN108564555A (en) * 2018-05-11 2018-09-21 中北大学 A kind of digital image noise reduction method based on NSST and CNN
CN109086806A (en) * 2018-07-16 2018-12-25 福州大学 A kind of IOT portable device visual identity accelerated method based on low resolution, compressed image
CN109492766A (en) * 2018-11-07 2019-03-19 西安交通大学 A kind of width learning method based on minimum P norm
CN109886971A (en) * 2019-01-24 2019-06-14 西安交通大学 A kind of image partition method and system based on convolutional neural networks

Non-Patent Citations (3)

* Cited by examiner, † Cited by third party
Title
JUNWEI JIN 等: "Discriminative graph regularized broad learning system for image recognition", 《SCIENCE CHINA（INFORMATION SCIENCES）》 *
李传朋等: "基于深度卷积神经网络的图像去噪研究", 《计算机工程》 *
贾晨 等: "基于宽度学习方法的多模态信息融合", 《智能系统学报》 *

Cited By (23)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110837890A (en) * 2019-10-22 2020-02-25 西安交通大学 Weight value fixed-point quantization method for lightweight convolutional neural network
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
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
CN110956202A (en) * 2019-11-13 2020-04-03 重庆大学 Image training method, system, medium and intelligent device based on distributed learning
CN110909796B (en) * 2019-11-22 2022-05-17 浪潮电子信息产业股份有限公司 Image classification method and related device
CN110909796A (en) * 2019-11-22 2020-03-24 浪潮电子信息产业股份有限公司 Image classification method and related device
CN110929652A (en) * 2019-11-26 2020-03-27 天津大学 Handwritten Chinese character recognition method based on LeNet-5 network model
CN110929652B (en) * 2019-11-26 2023-08-01 天津大学 Handwriting Chinese character recognition method based on LeNet-5 network model
CN112861896A (en) * 2019-11-27 2021-05-28 北京沃东天骏信息技术有限公司 Image identification method and device
CN111476138A (en) * 2020-03-31 2020-07-31 万翼科技有限公司 Construction method and identification method of building drawing component identification model and related equipment
CN111476138B (en) * 2020-03-31 2023-08-18 万翼科技有限公司 Construction method, identification method and related equipment for building drawing component identification model
CN113642592A (en) * 2020-04-27 2021-11-12 武汉Tcl集团工业研究院有限公司 Training method of training model, scene recognition method and computer equipment
CN111614358A (en) * 2020-04-30 2020-09-01 北京的卢深视科技有限公司 Method, system, device and storage medium for feature extraction based on sub-channel quantization
CN111614358B (en) * 2020-04-30 2023-08-04 合肥的卢深视科技有限公司 Feature extraction method, system, equipment and storage medium based on multichannel quantization
CN112070100A (en) * 2020-09-11 2020-12-11 深圳力维智联技术有限公司 Image feature recognition method and device based on deep learning model and storage medium
CN112906829A (en) * 2021-04-13 2021-06-04 成都四方伟业软件股份有限公司 Digital recognition model construction method and device based on Mnist data set
CN113205177B (en) * 2021-04-25 2022-03-25 广西大学 Electric power terminal identification method based on incremental collaborative attention mobile convolution
CN113205177A (en) * 2021-04-25 2021-08-03 广西大学 Electric power terminal identification method based on incremental collaborative attention mobile convolution
CN114363477A (en) * 2021-12-30 2022-04-15 上海网达软件股份有限公司 Method and system for video self-adaptive sharpening based on sliding window weight regression
CN114444622A (en) * 2022-04-11 2022-05-06 中国科学院微电子研究所 Fruit detection system and method based on neural network model
CN114444622B (en) * 2022-04-11 2022-06-17 中国科学院微电子研究所 Fruit detection system and method based on neural network model
CN114781650A (en) * 2022-04-28 2022-07-22 北京百度网讯科技有限公司 Data processing method, device, equipment and storage medium
CN114781650B (en) * 2022-04-28 2024-02-27 北京百度网讯科技有限公司 Data processing method, device, equipment and storage medium

Also Published As

Publication number Publication date
CN110288030B (en) 2023-04-07

Similar Documents

Publication Publication Date Title
CN110288030A (en) Image-recognizing method, device and equipment based on lightweight network model
CN110020682B (en) Attention mechanism relation comparison network model method based on small sample learning
US20210049423A1 (en) Efficient image classification method based on structured pruning
US20190228268A1 (en) Method and system for cell image segmentation using multi-stage convolutional neural networks
CN110348399B (en) Hyperspectral intelligent classification method based on prototype learning mechanism and multidimensional residual error network
Thai et al. Image classification using support vector machine and artificial neural network
CN107169504B (en) A kind of hand-written character recognition method based on extension Non-linear Kernel residual error network
CN103927531B (en) It is a kind of based on local binary and the face identification method of particle group optimizing BP neural network
CN109299741B (en) Network attack type identification method based on multi-layer detection
CN108614997B (en) Remote sensing image identification method based on improved AlexNet
CN103955702A (en) SAR image terrain classification method based on depth RBF network
CN111882040A (en) Convolutional neural network compression method based on channel number search
CN110197205A (en) A kind of image-recognizing method of multiple features source residual error network
CN104616029B (en) Data classification method and device
CN105243139A (en) Deep learning based three-dimensional model retrieval method and retrieval device thereof
CN105512681A (en) Method and system for acquiring target category picture
CN104318515B (en) High spectrum image wave band dimension reduction method based on NNIA evolution algorithms
CN112949738B (en) Multi-class unbalanced hyperspectral image classification method based on EECNN algorithm
CN103440493A (en) Hyperspectral image blur classification method and device based on related vector machine
CN112800927B (en) Butterfly image fine-granularity identification method based on AM-Softmax loss
Wang et al. Energy based competitive learning
CN107545279A (en) Image-recognizing method based on convolutional neural networks Yu Weighted Kernel signature analysis
Li et al. Dating ancient paintings of Mogao Grottoes using deeply learnt visual codes
CN112766283A (en) Two-phase flow pattern identification method based on multi-scale convolution network
CN111222545B (en) Image classification method based on linear programming incremental learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant