CN108921198A - commodity image classification method, server and system based on deep learning - Google Patents
commodity image classification method, server and system based on deep learning Download PDFInfo
- Publication number
- CN108921198A CN108921198A CN201810588841.7A CN201810588841A CN108921198A CN 108921198 A CN108921198 A CN 108921198A CN 201810588841 A CN201810588841 A CN 201810588841A CN 108921198 A CN108921198 A CN 108921198A
- Authority
- CN
- China
- Prior art keywords
- commodity image
- training
- classification
- commodity
- image
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/26—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion
- G06V10/267—Segmentation of patterns in the image field; Cutting or merging of image elements to establish the pattern region, e.g. clustering-based techniques; Detection of occlusion by performing operations on regions, e.g. growing, shrinking or watersheds
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses commodity image classification method, server and systems based on deep learning.Wherein this method, including obtain commodity image and be labeled classification, form commodity image classification based training data set and commodity image class test data set；Background segment is carried out to the commodity image that training data is concentrated using SVM algorithm；Training dataset after background segment is input in the convolutional neural networks of the image classification of pre-training completion, and carries out parameter adjustment using convolutional neural networks of the method for transfer learning to the image classification that pre-training is completed, constructs commodity image disaggregated model；Commodity image class test data set is input in commodity image disaggregated model and carries out prediction classification, until the prediction classification of commodity image disaggregated model output reaches default precision, commodity image disaggregated model is completed in final training；Commodity image to be sorted is input in the commodity image disaggregated model of training completion, output prediction classification results.
Description
Technical field
The invention belongs to image identification technical field more particularly to a kind of commodity image classification sides based on deep learning
Method, server and system.
Background technique
After the internet of things era arrives, e-commerce rapidly develops, and more and more people get used to using shopping at network.Net
Network shopping brings great convenience, and time whole world can be bought by staying indoors, nowadays the shopping at network platform of mainstream, such as Taobao,
Jingdone district, Amazon etc., most of is the product search system based on keyword.There is very big drawback in this mode：One side
Face, the product search system based on keyword require businessman to classify first to commodity image, then add corresponding text
Mark, still, these marks are difficult to reflect product features comprehensively, simultaneously because commodity image quantity sharply increases, it is artificial to mark
Note takes time and effort, and greatly reduces the working efficiency of businessman.On the other hand, user inputs crucial according to the demand for commodity of oneself
Subjectivity is had when word scans for, there is different user demands and but has input identical keyword in this meeting maximum probability, or
It is that identical user demand has input different keywords, both of which will lead to the commodity image of retrieved web appearance
As a result inconsistent with user's expectation, it also will be greatly reduced the efficiency that user buys expectation commodity.How shopping website is improved
Commodity recall precision the problem of having become people's extensive concern, being scanned on website using image data can be with
Retrieval is reduced for the dependence of text, is a solution for being worth research.
With the development of electronics technology, these defects can greatly be made up by electronic information technology, be based on commodity image
The image procossing mode of content is come into being, and in this way, can intuitively show the information of commodity, without artificial
It is labeled, simple and efficient, application is gradually extensive.But meanwhile traditional image classification method based on feature extracts spy
Sign is complicated, and single features can not obtain preferable classification accuracy.
In conclusion how to solve the problems, such as the Accurate classification of shiploads of merchandise image, still lack effective solution scheme.
Summary of the invention
In order to solve the deficiencies in the prior art, the first object of the present invention is to provide a kind of commodity figure based on deep learning
As classification method, can rapidly and accurately classify to shiploads of merchandise image.
A kind of commodity image classification method based on deep learning of the invention, including：
It obtains commodity image and is labeled classification, form commodity image classification based training data set and commodity image classification is surveyed
Try data set；
Background segment is carried out to the commodity image in commodity image classification based training data set using SVM algorithm；
Commodity image classification based training data set after background segment is input to the convolution of the image classification of pre-training completion
In neural network, and parameter is carried out using convolutional neural networks of the method for transfer learning to the image classification that pre-training is completed
Adjustment, constructs commodity image disaggregated model；
Commodity image class test data set is input in commodity image disaggregated model and carries out prediction classification, until commodity
The prediction classification of image classification model output reaches default precision, and commodity image disaggregated model is completed in final training；
Commodity image to be sorted is input in the commodity image disaggregated model of training completion, output prediction classification knot
Fruit.
Further, background segment is carried out to the commodity image in commodity image classification based training data set using SVM algorithm
Process, including：
Commodity image to be split is done into prospect label and background label；
Several pixels and extract individual features as training sample from prospect label and background label respectively；
The parameter and kernel function for determining SVM classifier, using the individual features training SVM classifier of extraction；
Classified using the SVM classifier that training is completed to the pixel of commodity image to be split, realizes commodity image
Background segment.
Further, the Selection of kernel function Polynomial kernel function of SVM classifier.
Further, the convolutional neural networks for the image classification completed in the method using transfer learning to pre-training carry out
During parameter adjusts, carried out according to convolutional neural networks of the convolution kernel capacity principle to the image classification that pre-training is completed
It improves.
The second object of the present invention is a kind of commodity image classified service device based on deep learning.
A kind of commodity image classified service device based on deep learning of the invention, including：
Data set acquisition module is configured as：It obtains commodity image and is labeled classification, form commodity image classification
Training dataset and commodity image class test data set；
Background segment module, is configured as：Using SVM algorithm to the commodity figure in commodity image classification based training data set
As carrying out background segment；
Commodity image disaggregated model constructs module, is configured as：By the commodity image classification based training number after background segment
It is input to according to collection in the convolutional neural networks of the image classification of pre-training completion, and using the method for transfer learning to pre-training
The convolutional neural networks of the image classification of completion carry out parameter adjustment, construct commodity image disaggregated model；
Commodity image disaggregated model test module, is configured as：Commodity image class test data set is input to quotient
Prediction classification is carried out in product image classification model, until the prediction classification of commodity image disaggregated model output reaches default precision,
Commodity image disaggregated model is completed in final training；
It predicts categorization module, is configured as：Commodity image to be sorted is input to the commodity image point of training completion
In class model, output prediction classification results.
Further, the background segment module, including：
Label setup module, is configured as：Commodity image to be split is done into prospect label and background label；
Characteristic extracting module is configured as：Several pixels are used as instruction from prospect label and background label respectively
Practice sample, and extracts individual features；
SVM classifier training module, is configured as：The parameter and kernel function for determining SVM classifier, using the phase of extraction
Answer feature training SVM classifier；
SVM classifier categorization module, is configured as：The SVM classifier completed using training is to commodity image to be split
Pixel classify, realize the background segment of commodity image.
Further, in the SVM classifier training module, the Selection of kernel function Polynomial kernel function of SVM classifier.
Further, in commodity image disaggregated model building module, according to convolution kernel capacity principle to pre-training
The convolutional neural networks of the image classification of completion improve.
The third object of the present invention is to provide a kind of commodity image categorizing system based on deep learning.
A kind of commodity image categorizing system based on deep learning of the invention, including described above based on deep learning
Commodity image classified service device.
Compared with prior art, the beneficial effects of the invention are as follows：
(1) present invention employs SVM classifiers has carried out background segment to commodity image, classifies compared to traditional images,
After being split to commodity image background, the accuracy rate of classification can be improved.
(2) present invention has also carried out pre-training to network structure, and the parameter after pre-training moves to conduct in new network
Commodity image class test data set is input in commodity image disaggregated model and carries out prediction classification by initiation parameter, until
The prediction classification of commodity image disaggregated model output reaches default precision, and commodity image disaggregated model is completed in final training, improves
The accuracy rate of commodity image classification.
Detailed description of the invention
The accompanying drawings constituting a part of this application is used to provide further understanding of the present application, and the application's shows
Meaning property embodiment and its explanation are not constituted an undue limitation on the present application for explaining the application.
Fig. 1 is a kind of commodity image classification method flow chart based on deep learning of the invention.
Fig. 2 is space reflection schematic diagram.
Fig. 3 (a) is the commodity image before segmentation.
Fig. 3 (b) is the image after SVM background segment.
Fig. 4 is the convolutional neural networks structural schematic diagram for the image classification that the pre-training that the present invention uses is completed.
Fig. 5 is convolutional layer schematic diagram.
Fig. 6 (a) is Sigmoid function.
Fig. 6 (b) is ReLU activation primitive.
Fig. 7 is AlexNet structural schematic diagram.
Fig. 8 is a kind of commodity image classified service device structural schematic diagram based on deep learning of the invention.
Specific embodiment
It is noted that following detailed description is all illustrative, it is intended to provide further instruction to the application.Unless another
It indicates, all technical and scientific terms used herein has logical with the application person of an ordinary skill in the technical field
The identical meanings understood.
It should be noted that term used herein above is merely to describe specific embodiment, and be not intended to restricted root
According to the illustrative embodiments of the application.As used herein, unless the context clearly indicates otherwise, otherwise singular shape
Formula be also intended to include plural form, additionally, it should be understood that, when in the present specification use term "comprising" and/or
When " comprising ", existing characteristics, step, operation, device, component and/or their combination are indicated.
Fig. 1 is a kind of commodity image classification method flow chart based on deep learning of the invention.
As shown in Figure 1, a kind of commodity image classification method based on deep learning of the invention, including：
Step 1：It obtains commodity image and is labeled classification, form commodity image classification based training data set and commodity image
Class test data set.
Specifically, the data set of the present embodiment uses more extensive Taobao day cat, Jingdone district, Amazon etc. from current
Major shopping website includes 10 class commodity altogether, and part of image has background, and image size is 256*256, picture format
For .jpg.
Step 2：Background segment is carried out to the commodity image in commodity image classification based training data set using SVM algorithm.
Using SVM algorithm, experiment porch is based on matlab for the segmentation of the commodity image background of the present embodiment
Libsvm under 2016a.
SVM algorithm has excellent generalization ability and recognition capability, in terms of image background segmentation, has good effect
Fruit.SVM is that input variable is mapped to a high-dimensional feature space by the Nonlinear Mapping selected in advance by certain, at this
Optimal separating hyper plane is constructed in space.The process that data set classification work is carried out using SVM is with previously selected one first
The input space is mapped to high-dimensional feature space as shown in Fig. 2, by being mapped to high-dimensional feature space by a little Nonlinear Mappings,
The nonlinear data for being not easy to be partitioned into itself in plane, which has been divided, to come.
This Nonlinear Mapping is realized by kernel function, and common kernel function includes Polynomial kernel function, Gauss
Core (RBF core) function and linear kernel function.
Although the value of kernel function is that it is the conversion carried out feature from low-dimensional to higher-dimension, kernel function is in advance low
Calculated in dimension, and substantial classifying quality shown on higher-dimension, also avoid directly in higher dimensional space into
The complicated calculating of row.
Wherein, the mistake of background segment is carried out to the commodity image in commodity image classification based training data set using SVM algorithm
Journey, including：
Step 2.1：Commodity image to be split is done into prospect label and background label；
Step 2.2：Several pixels and extract corresponding as training sample from prospect label and background label respectively
Feature；
Step 2.3：The parameter and kernel function for determining SVM classifier, using the individual features training SVM classifier of extraction；
Step 2.4：Classified using the SVM classifier that training is completed to the pixel of commodity image to be split, is realized
The background segment of commodity image.
Wherein, the Selection of kernel function Polynomial kernel function of SVM classifier.
When input picture is the commodity image that band has powerful connections in data set, in SVM algorithm setting, the back of image to be split
Scape is labeled as 0, and prospect is labeled as 1, and foreground and background respectively has chosen 20 points as training sample, makes when pixel is chosen
It is taken a little with the help of ginput function.Selection of kernel function Polynomial kernel function, degree are set as -1.Experimental result such as Fig. 3
(a) and shown in Fig. 3 (b), Fig. 3 (a) is the commodity image before segmentation, and Fig. 3 (b) is the image after SVM background segment.From experiment
As a result as can be seen that SVM algorithm is well divided the background of image in, and profile information saves well.
Step 3：Commodity image classification based training data set after background segment is input to the image classification of pre-training completion
Convolutional neural networks in, and the convolutional neural networks for image classification pre-training completed using the method for transfer learning into
The adjustment of row parameter, constructs commodity image disaggregated model.
Present invention employs the transfer learnings based on model, while occurring negative transfer in order to prevent, used in this method
The data set and commodity image of pre-training have enough relevances, therefore can carry out relevant transfer learning.
Specifically, the convolutional neural networks for the image classification that the pre-training that the present invention uses is completed, as shown in figure 4, by rolling up
Lamination, pond layer, full articulamentum composition.
Convolutional layer：In the application of image procossing, the effect of convolutional layer is to extract characteristics of image.Convolutional layer contains many volumes
Product core, each convolution kernel can obtain a kind of feature to image convolution.
As shown in figure 5, black box is convolution kernel, it is assumed herein that step-length is set as 1, convolution operation is expressed as moving right every time
A pixel is moved, then sequentially moves down a pixel again.Number is marked in cell to indicate the power of each convolution kernel
Weight, such as red have 4 weights.Picture pixels and the corresponding power of convolution kernel in convolution process, that is, convolution kernel moving process
Multiplied by weight is carried out again, is finally added and is obtained an output.Fortune in this way, passes through the continuous convolution of convolutional layer, image institute
The feature for including is extracted.
Pond layer：Pond layer i.e. down-sampling layer, main function is that a dimensionality reduction operation is carried out to characteristic pattern, to have
Effect reduces the parameter that subsequent neural net layer needs, while greatly reducing calculation amount.Chi Huayou is averaged pondization and maximum pond
Deng key property has：Invariance, including invariable rotary, Scale invariant etc., focus can be placed on characteristics of image by this
In itself and on position present in non-image features；Export fixed length, i.e., by pond layer after, the size of characteristic pattern can subtract
The small half for former characteristic pattern；Network parameter can be reduced while retaining main feature, reduced calculation amount and prevented simultaneously
Over-fitting greatly improves the generalization ability of network.
Full articulamentum：The effect of full articulamentum be the two dimensional character of convolutional layer final output is converted to finally it is one-dimensional
Vector.The core operation connected entirely is exactly linear change of the matrix-vector product by a feature space to another feature space
It changes, each tie point is connected with all tie points of preceding layer, and effect is the feature that comprehensive prior process is extracted.
The input of neuron in full articulamentum is indicated using x, a indicates the output of neuron, and b is biasing, wherein under each
Mark respectively indicates i-th of relevant parameter (i=1,2,3).
wijIndicate between i-th of neuron and j-th of neuron connection weight (i=1,2,3；J=1,2,3).
The matrix representation forms of articulamentum are entirely：
Activation primitive：In neural network, signal received by each neuron is the line of previous neuron
Property function, therefore, in order to avoid linear coincidence, it is that each neuron introduces non-linear factor that activation primitive, which can be used,.This
It is that complete expressiveness is not had due to linear model, so introducing activation primitive.Convolutional neural networks commonly activate
Function has ReLU function and Sigmoid function.Wherein the expression formula of Sigmoid function is：
Wherein, x indicates the input of neuron in full articulamentum.
As shown in Fig. 6 (a), Sigmoid function has a big signal gain in central area, and two sides have smaller signal
Gain, it has good effect to the feature space figure of signal, but there is also certain limitations simultaneously：One, when far from origin
Functional gradient levels off to 0, will lead to training process and gradient disappearance occurs；Two, function output reduces not centered on origin
Weight updates efficiency；Three, a large amount of exponent arithmetic is introduced, calculation amount is increased, reduces arithmetic speed.
The expression formula of ReLU function is：
F (x)=max (0, x)
Wherein, x indicates the input of neuron in full articulamentum.
ReLU has been proved to it effectively can inhibit gradient to disappear compared to traditional activation primitive, while also have more
Fast convergence rate, therefore the application of ReLU activation primitive is comparatively more extensive.Shown in its image such as Fig. 6 (b).
Dropout：Dropout is proposed for the over-fitting occurred for network training process kind, and Dropout can
It is not understood as in the training process of neural network, random abandons some units to simplify the process of network.By
The link information of neural network random drop a part after Dropout operation.When the numerical value of Dropout is set as 0.5
Have the effect of it is best, the reason is that most network structures can be generated in Dropout when numerical value is set as 0.5.Dropout is random
Hidden node is ignored in selection, is different from due to ignoring at random every time, this results in the network trained every time all different,
Therefore each training process can be regarded as to the model for establishing one " new " completely.
Caffe frame：With the rise of deep learning upsurge, deep learning frame is continued to bring out, and is deep learning network
Realization provide a great convenience, for designer built basis function realize, save the plenty of time.Wherein Caffe
Deep learning frame be one efficiently, clearly deep learning frame, it has the features such as modularization, Test coverage, and simultaneously
Two kinds of interfaces of Python and Matlab are provided, and can realize the switching between CPU and GPU by function call, in order to just
True committed memory.Caffe follow neural network it is assumed that i.e. all calculating all indicates in the form of layer, each layer
Function is exactly to obtain some data, the result after then output calculates.Wherein each layer does two calculating：Propagated forward
That output is calculated from input, backpropagation be from above to gradient calculate the gradient relative to input, the two calculate real
It after now, allows for plurality of layers connection and forms a network, in this way, whole network is exactly input data (image), then come
Calculate the output (label) needed.When training, loss function and gradient can be calculated according to existing label, so
Afterwards with gradient come backpropagation, the parameter of corrective networks, here it is the basic procedures of Caffe.
SIFT and HOG feature：SIFT feature, that is, Scale invariant features transform is to detect key point in the picture, is one
Kind local feature description, has scale invariability；The local grain that histograms of oriented gradients (HOG) is used to describe image is special
Sign.The advantage of HOG feature extraction is that it covers texture information abundant, while introducing this physical quantity of amplitude, therefore
It can make rotation and translation that there is invariance.But it is more sensitive in the feature extraction to noise spot, and be easy by outer
The influence of portion's environment.
In order to facilitate the migration of parameter, network structure of the invention is modified from AlexNet, as shown in Figure 7.
Wherein, AlexNet is Hinton group neural network model used in ISVRC2012.Network structure is not
It is that the label layer of the last layer is modified by 1000 classes for 10 classes with place.The input of network is the colored RGB of 256*256
The unified size of image can be cut to 227*227 and facilitate convolution operation by image, data Layer.Convolution kernel size point in network
Not Wei 11*11,5*5 and 3*3, pond mode is maximum pond, and Dropout is set as 0.5.It can be generated after convolution operation big
The size of small different characteristic pattern, each layer characteristic pattern is 55*55,27*27,27*27,13*13,13*13,13*13 and 6* respectively
6, the number of characteristic pattern is respectively 96,96,256,256,384,384,256 and 256.The parameter of full articulamentum is set as 4096,
The classification prediction of the last layer uses softmax classifier.
Specifically, the convolutional neural networks for the image classification that pre-training is completed are joined using the method for transfer learning
During number adjustment, the convolutional neural networks for the image classification that pre-training is completed are changed according to convolution kernel capacity principle
Into.
The effect of convolution kernel is to extract characteristics of image to find out picture structure, and convolution kernel capacity is one and is used to measure convolution
Core finds out the scale of the ability of picture structure.If convolution kernel capacity is small, it is meant that can only have local feature mapping in image
To next layer, but if convolution kernel capacity is big, it is meant that can have more Feature Mappings to next layer network, work as convolution
When the value of core capacity is greater than 1/6, convolution kernel has good detectivity.
The calculation formula of convolution kernel capacity is：
If the size of convolution kernel is n × n, operated by the maximum pondization of a 3*3, then this convolution kernel is true
Convolution kernel is sized for 3n × 3n, if by the maximum pond of m 3*3, then the actual size of convolution kernel is just 3mn×
3mn。
Receptive field is an important concept in convolutional neural networks, it indicates a specific convolution in the input space
The range areas of neural network feature.The size of a general receptive field can use feature sizes and center in region
It is described.In the operation of the convolution of neural network：The size of current receptive field is r, and the distance between adjacent feature is j,
Convolution kernel is filled with p, step-length s having a size of p.
So export the size n of featureoutWith input feature vector size ninRelationship be represented by：
Export the distance j between characteristic patternoutDistance j between input feature vector figureinRelationship be：
jout=jin*s
The receptive field r of outputoutWith the impression r of inputinWild relationship is：
rout=rin+(k-1)*jin
According to the receptive field size and convolution kernel capacity of each convolutional layer of network structure that above-mentioned calculated relationship acquires the design,
The convolution kernel capacity of improved network is generally higher than AlexNet, that is to say, that the convolution kernel of network has stronger spy after improvement
Extractability is levied, this is also the major reason that network obtains preferable classifying quality after improving.
Step 4：Commodity image class test data set is input in commodity image disaggregated model and carries out prediction classification, directly
Reach default precision to the prediction classification that commodity image disaggregated model exports, commodity image disaggregated model is completed in final training.
Step 5：Commodity image to be sorted is input in the commodity image disaggregated model of training completion, output prediction point
Class result.
Experimental situation：
The related hardware for testing computer used is configured to：Intel(R)Xeon(R)CPU E5-2620@2.10GHz×
15, physical memory is：16.00GB.GPU be two pieces of NAVIDIA TITAN X, 1.53GHz dominant frequency, 12GB GDDR5X video memory,
12000000000 number of transistors, 3584 CUDA cores.Experimental implementation system is Ubuntu 14.04, and distribution experiment operation platform is base
In the deep learning frame caffe of C++, configuration interface is python interface.
Fig. 8 is a kind of commodity image classified service device structural schematic diagram based on deep learning of the invention.
As shown in figure 8, a kind of commodity image classified service device based on deep learning of the invention, including：
(1) data set acquisition module is configured as：It obtains commodity image and is labeled classification, form commodity image
Classification based training data set and commodity image class test data set；
(2) background segment module is configured as：Using SVM algorithm to the quotient in commodity image classification based training data set
Product image carries out background segment；
Specifically, the background segment module, including：
Label setup module, is configured as：Commodity image to be split is done into prospect label and background label；
Characteristic extracting module is configured as：Several pixels are used as instruction from prospect label and background label respectively
Practice sample, and extracts individual features；
SVM classifier training module, is configured as：The parameter and kernel function for determining SVM classifier, using the phase of extraction
Answer feature training SVM classifier；
SVM classifier categorization module, is configured as：The SVM classifier completed using training is to commodity image to be split
Pixel classify, realize the background segment of commodity image.
Specifically, in the SVM classifier training module, the Selection of kernel function Polynomial kernel function of SVM classifier.
(3) commodity image disaggregated model constructs module, is configured as：By the commodity image classification based training after background segment
Data set is input in the convolutional neural networks of the image classification of pre-training completion, and using the method for transfer learning to pre- instruction
The convolutional neural networks for practicing the image classification completed carry out parameter adjustment, construct commodity image disaggregated model；
Specifically, complete to pre-training according to convolution kernel capacity principle in commodity image disaggregated model building module
At the convolutional neural networks of image classification improve.
(4) commodity image disaggregated model test module, is configured as：Commodity image class test data set is input to
Prediction classification is carried out in commodity image disaggregated model, until the prediction classification of commodity image disaggregated model output reaches default essence
Commodity image disaggregated model is completed in degree, final training；
(5) it predicts categorization module, is configured as：Commodity image to be sorted is input to the commodity image of training completion
In disaggregated model, output prediction classification results.
The commodity image categorizing system based on deep learning that the present invention also provides a kind of.
A kind of commodity image categorizing system based on deep learning of the invention, including as shown in Figure 8 based on depth
The commodity image classified service device of habit.
Present invention employs SVM classifiers to have carried out background segment to commodity image, classifies compared to traditional images, right
After commodity image background is split, the accuracy rate of classification can be improved.
The present invention has also carried out pre-training to network structure, and the parameter after pre-training moves in new network as initial
Change parameter, commodity image class test data set is input in commodity image disaggregated model and carries out prediction classification, until commodity
The prediction classification of image classification model output reaches default precision, and commodity image disaggregated model is completed in final training, improves quotient
The accuracy rate of product image classification.
It should be understood by those skilled in the art that, the embodiment of the present invention can provide as method, system or computer program
Product.Therefore, the shape of hardware embodiment, software implementation or embodiment combining software and hardware aspects can be used in the present invention
Formula.Moreover, the present invention, which can be used, can use storage in the computer that one or more wherein includes computer usable program code
The form for the computer program product implemented on medium (including but not limited to magnetic disk storage and optical memory etc.).
The present invention be referring to according to the method for the embodiment of the present invention, the process of equipment (system) and computer program product
Figure and/or block diagram describe.It should be understood that can be realized by computer program instructions each in flowchart and/or the block diagram
The combination of process and/or box in process and/or box and flowchart and/or the block diagram.It can provide these computers
Processor of the program instruction to general purpose computer, special purpose computer, Embedded Processor or other programmable data processing devices
To generate a machine, so that being generated by the instruction that computer or the processor of other programmable data processing devices execute
For realizing the function of being specified in one or more flows of the flowchart and/or one or more blocks of the block diagram
Device.
These computer program instructions, which may also be stored in, is able to guide computer or other programmable data processing devices with spy
Determine in the computer-readable memory that mode works, so that instruction stored in the computer readable memory generation includes
The manufacture of command device, the command device are realized in one box of one or more flows of the flowchart and/or block diagram
Or the function of being specified in multiple boxes.
These computer program instructions also can be loaded onto a computer or other programmable data processing device, so that counting
Series of operation steps are executed on calculation machine or other programmable devices to generate computer implemented processing, thus in computer
Or the instruction executed on other programmable devices is provided for realizing in one or more flows of the flowchart and/or box
The step of function of being specified in figure one box or multiple boxes.
Those of ordinary skill in the art will appreciate that realizing all or part of the process in above-described embodiment method, being can be with
Relevant hardware is instructed to complete by computer program, the program can be stored in a computer-readable storage and be situated between
In matter, the program is when being executed, it may include such as the process of the embodiment of above-mentioned each method.Wherein, the storage medium can be
Magnetic disk, CD, read-only memory (Read-Only Memory, ROM) or random access memory (Random
AccessMemory, RAM) etc..
Above-mentioned, although the foregoing specific embodiments of the present invention is described with reference to the accompanying drawings, not protects model to the present invention
The limitation enclosed, those skilled in the art should understand that, based on the technical solutions of the present invention, those skilled in the art
The various modifications or changes that can be made are not needed to make the creative labor still within protection scope of the present invention.
Claims (9)
1. a kind of commodity image classification method based on deep learning, which is characterized in that including：
It obtains commodity image and is labeled classification, form commodity image classification based training data set and commodity image class test number
According to collection；
Background segment is carried out to the commodity image in commodity image classification based training data set using SVM algorithm；
Commodity image classification based training data set after background segment is input to the convolutional Neural of the image classification of pre-training completion
In network, and parameter adjustment is carried out using convolutional neural networks of the method for transfer learning to the image classification that pre-training is completed,
Construct commodity image disaggregated model；
Commodity image class test data set is input in commodity image disaggregated model and carries out prediction classification, until commodity image
The prediction classification of disaggregated model output reaches default precision, and commodity image disaggregated model is completed in final training；
Commodity image to be sorted is input in the commodity image disaggregated model of training completion, output prediction classification results.
2. a kind of commodity image classification method based on deep learning as described in claim 1, which is characterized in that use SVM
Algorithm carries out the process of background segment to the commodity image in commodity image classification based training data set, including：
Commodity image to be split is done into prospect label and background label；
Several pixels and extract individual features as training sample from prospect label and background label respectively；
The parameter and kernel function for determining SVM classifier, using the individual features training SVM classifier of extraction；
Classified using the SVM classifier that training is completed to the pixel of commodity image to be split, realizes the back of commodity image
Scape segmentation.
3. a kind of commodity image classification method based on deep learning as claimed in claim 2, which is characterized in that svm classifier
The Selection of kernel function Polynomial kernel function of device.
4. a kind of commodity image classification method based on deep learning as described in claim 1, which is characterized in that moved using
During the method for moving study carries out parameter adjustment to the convolutional neural networks for the image classification that pre-training is completed, according to convolution
Core capacity principle improves the convolutional neural networks for the image classification that pre-training is completed.
5. a kind of commodity image classified service device based on deep learning, which is characterized in that including：
Data set acquisition module is configured as：It obtains commodity image and is labeled classification, form commodity image classification based training
Data set and commodity image class test data set；
Background segment module, is configured as：Using SVM algorithm to the commodity image in commodity image classification based training data set into
Row background segment；
Commodity image disaggregated model constructs module, is configured as：By the commodity image classification based training data set after background segment
It is input in the convolutional neural networks of the image classification of pre-training completion, and pre-training is completed using the method for transfer learning
The convolutional neural networks of image classification carry out parameter adjustment, construct commodity image disaggregated model；
Commodity image disaggregated model test module, is configured as：Commodity image class test data set is input to commodity figure
As carrying out prediction classification in disaggregated model, until the prediction classification of commodity image disaggregated model output reaches default precision, finally
Commodity image disaggregated model is completed in training；
It predicts categorization module, is configured as：Commodity image to be sorted is input to the commodity image classification mould of training completion
In type, output prediction classification results.
6. a kind of commodity image classified service device based on deep learning as claimed in claim 5, which is characterized in that the back
Scape divides module, including：
Label setup module, is configured as：Commodity image to be split is done into prospect label and background label；
Characteristic extracting module is configured as：Several pixels are used as training sample from prospect label and background label respectively
This, and extract individual features；
SVM classifier training module, is configured as：The parameter and kernel function for determining SVM classifier, using the corresponding spy of extraction
Levy training SVM classifier；
SVM classifier categorization module, is configured as：Using the trained SVM classifier completed to the picture of commodity image to be split
Vegetarian refreshments is classified, and realizes the background segment of commodity image.
7. a kind of commodity image classified service device based on deep learning as claimed in claim 6, which is characterized in that described
In SVM classifier training module, the Selection of kernel function Polynomial kernel function of SVM classifier.
8. a kind of commodity image classified service device based on deep learning as claimed in claim 5, which is characterized in that described
Commodity image disaggregated model constructs in module, the convolutional Neural according to the image classification that convolution kernel capacity principle completes pre-training
Network improves.
9. a kind of commodity image categorizing system based on deep learning, which is characterized in that including any one of such as claim 5-8
The commodity image classified service device based on deep learning.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810588841.7A CN108921198A (en) | 2018-06-08 | 2018-06-08 | commodity image classification method, server and system based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810588841.7A CN108921198A (en) | 2018-06-08 | 2018-06-08 | commodity image classification method, server and system based on deep learning |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108921198A true CN108921198A (en) | 2018-11-30 |
Family
ID=64420032
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810588841.7A Pending CN108921198A (en) | 2018-06-08 | 2018-06-08 | commodity image classification method, server and system based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108921198A (en) |
Cited By (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110059658A (en) * | 2019-04-26 | 2019-07-26 | 北京理工大学 | A kind of satellite-remote-sensing image multidate change detecting method based on Three dimensional convolution neural network |
| CN110348293A (en) * | 2019-05-29 | 2019-10-18 | 广州中科凯泽科技有限公司 | A kind of commodity recognition method and system |
| CN110458096A (en) * | 2019-08-09 | 2019-11-15 | 广州众聚智能科技有限公司 | A kind of extensive commodity recognition method based on deep learning |
| CN111082992A (en) * | 2019-12-23 | 2020-04-28 | 超讯通信股份有限公司 | SDN network data packet identification method based on deep learning |
| CN111191723A (en) * | 2019-12-30 | 2020-05-22 | 创新奇智(北京)科技有限公司 | Few-sample commodity classification system and method based on cascade classifier |
| CN111325115A (en) * | 2020-02-05 | 2020-06-23 | 山东师范大学 | Countermeasures cross-modal pedestrian re-identification method and system with triple constraint loss |
| CN111750965A (en) * | 2019-03-27 | 2020-10-09 | 杭州海康威视数字技术股份有限公司 | Commodity self-service charging method, device and system |
| CN112164076A (en) * | 2020-09-24 | 2021-01-01 | 济南大学 | Hardened cement water cement ratio prediction method and system based on cement microstructure image |
| CN112257646A (en) * | 2020-11-02 | 2021-01-22 | 创新奇智(南京)科技有限公司 | Commodity detection method and device, electronic equipment and storage medium |
| CN112801578A (en) * | 2021-01-18 | 2021-05-14 | 哈尔滨工业大学 | Commodity warehousing and ex-warehouse management system and method applied to individual vendor |
| CN113191378A (en) * | 2020-01-14 | 2021-07-30 | 阿里巴巴集团控股有限公司 | Prediction method, device and system for image search |
| CN113220970A (en) * | 2021-03-09 | 2021-08-06 | 十堰时风达工贸有限公司 | E-commerce big data platform based on block chain |
| CN113220918A (en) * | 2021-03-09 | 2021-08-06 | 十堰时风达工贸有限公司 | E-commerce big data classification method and system based on block technology |
| CN113361584A (en) * | 2021-06-01 | 2021-09-07 | 推想医疗科技股份有限公司 | Model training method and device, and pulmonary arterial hypertension measurement method and device |
| CN116579722A (en) * | 2023-07-14 | 2023-08-11 | 四川集鲜数智供应链科技有限公司 | Commodity distribution warehouse-in and warehouse-out management method based on deep learning |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103186901A (en) * | 2013-03-29 | 2013-07-03 | 中国人民解放军第三军医大学 | Full-automatic image segmentation method |
| CN106530283A (en) * | 2016-10-20 | 2017-03-22 | 北京工业大学 | SVM (support vector machine)-based medical image blood vessel recognition method |
| CN107239802A (en) * | 2017-06-28 | 2017-10-10 | 广东工业大学 | A kind of image classification method and device |
-
2018
- 2018-06-08 CN CN201810588841.7A patent/CN108921198A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103186901A (en) * | 2013-03-29 | 2013-07-03 | 中国人民解放军第三军医大学 | Full-automatic image segmentation method |
| CN106530283A (en) * | 2016-10-20 | 2017-03-22 | 北京工业大学 | SVM (support vector machine)-based medical image blood vessel recognition method |
| CN107239802A (en) * | 2017-06-28 | 2017-10-10 | 广东工业大学 | A kind of image classification method and device |
Non-Patent Citations (1)
| Title |
|---|
| 孙昂 等: "《基于改进的卷积神经网络多类商品精细分类》", 《山东师范大学学报(自然科学版)》 * |
Cited By (22)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111750965A (en) * | 2019-03-27 | 2020-10-09 | 杭州海康威视数字技术股份有限公司 | Commodity self-service charging method, device and system |
| CN110059658B (en) * | 2019-04-26 | 2020-11-24 | 北京理工大学 | Remote sensing satellite image multi-temporal change detection method based on three-dimensional convolutional neural network |
| CN110059658A (en) * | 2019-04-26 | 2019-07-26 | 北京理工大学 | A kind of satellite-remote-sensing image multidate change detecting method based on Three dimensional convolution neural network |
| CN110348293A (en) * | 2019-05-29 | 2019-10-18 | 广州中科凯泽科技有限公司 | A kind of commodity recognition method and system |
| CN110348293B (en) * | 2019-05-29 | 2022-11-18 | 广州中科凯泽科技有限公司 | Commodity identification method and system |
| CN110458096A (en) * | 2019-08-09 | 2019-11-15 | 广州众聚智能科技有限公司 | A kind of extensive commodity recognition method based on deep learning |
| CN111082992A (en) * | 2019-12-23 | 2020-04-28 | 超讯通信股份有限公司 | SDN network data packet identification method based on deep learning |
| CN111191723A (en) * | 2019-12-30 | 2020-05-22 | 创新奇智(北京)科技有限公司 | Few-sample commodity classification system and method based on cascade classifier |
| CN111191723B (en) * | 2019-12-30 | 2023-06-20 | 创新奇智(北京)科技有限公司 | Cascade classifier-based few-sample commodity classification system and classification method |
| CN113191378A (en) * | 2020-01-14 | 2021-07-30 | 阿里巴巴集团控股有限公司 | Prediction method, device and system for image search |
| CN111325115B (en) * | 2020-02-05 | 2022-06-21 | 山东师范大学 | Cross-modal countervailing pedestrian re-identification method and system with triple constraint loss |
| CN111325115A (en) * | 2020-02-05 | 2020-06-23 | 山东师范大学 | Countermeasures cross-modal pedestrian re-identification method and system with triple constraint loss |
| CN112164076B (en) * | 2020-09-24 | 2022-10-25 | 济南大学 | Hardened cement water cement ratio prediction method and system based on cement microstructure image |
| CN112164076A (en) * | 2020-09-24 | 2021-01-01 | 济南大学 | Hardened cement water cement ratio prediction method and system based on cement microstructure image |
| CN112257646A (en) * | 2020-11-02 | 2021-01-22 | 创新奇智(南京)科技有限公司 | Commodity detection method and device, electronic equipment and storage medium |
| CN112257646B (en) * | 2020-11-02 | 2023-09-12 | 创新奇智(南京)科技有限公司 | Commodity detection method and device, electronic equipment and storage medium |
| CN112801578A (en) * | 2021-01-18 | 2021-05-14 | 哈尔滨工业大学 | Commodity warehousing and ex-warehouse management system and method applied to individual vendor |
| CN113220970A (en) * | 2021-03-09 | 2021-08-06 | 十堰时风达工贸有限公司 | E-commerce big data platform based on block chain |
| CN113220918A (en) * | 2021-03-09 | 2021-08-06 | 十堰时风达工贸有限公司 | E-commerce big data classification method and system based on block technology |
| CN113361584A (en) * | 2021-06-01 | 2021-09-07 | 推想医疗科技股份有限公司 | Model training method and device, and pulmonary arterial hypertension measurement method and device |
| CN116579722A (en) * | 2023-07-14 | 2023-08-11 | 四川集鲜数智供应链科技有限公司 | Commodity distribution warehouse-in and warehouse-out management method based on deep learning |
| CN116579722B (en) * | 2023-07-14 | 2023-09-19 | 四川集鲜数智供应链科技有限公司 | Commodity distribution warehouse-in and warehouse-out management method based on deep learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108921198A (en) | commodity image classification method, server and system based on deep learning | |
| CN110866140B (en) | Image feature extraction model training method, image searching method and computer equipment | |
| CN112733749B (en) | Real-time pedestrian detection method integrating attention mechanism | |
| CN112396002B (en) | SE-YOLOv 3-based lightweight remote sensing target detection method | |
| Tang et al. | Deepchart: Combining deep convolutional networks and deep belief networks in chart classification | |
| Liu et al. | ARC-Net: An efficient network for building extraction from high-resolution aerial images | |
| Quoc Bao et al. | Plant species identification from leaf patterns using histogram of oriented gradients feature space and convolution neural networks | |
| CN108875076B (en) | Rapid trademark image retrieval method based on Attention mechanism and convolutional neural network | |
| Sun et al. | Fast object detection based on binary deep convolution neural networks | |
| CN109063719B (en) | Image classification method combining structure similarity and class information | |
| CN109117781A (en) | Method for building up, device and the more attribute recognition approaches of more attribute Recognition Models | |
| CN110633708A (en) | Deep network significance detection method based on global model and local optimization | |
| Ye et al. | A new re-ranking method based on convolutional neural network and two image-to-class distances for remote sensing image retrieval | |
| CN113807399A (en) | Neural network training method, neural network detection method and neural network detection device | |
| Xiong et al. | MSN: Modality separation networks for RGB-D scene recognition | |
| Cui et al. | Lightweight spectral–spatial attention network for hyperspectral image classification | |
| Hashemi Hosseinabad et al. | Multiple answers to a question: a new approach for visual question answering | |
| Li et al. | Page segmentation using convolutional neural network and graphical model | |
| Wei et al. | Food image classification and image retrieval based on visual features and machine learning | |
| Xin et al. | Hybrid dilated multilayer faster RCNN for object detection | |
| Chen et al. | Research on object detection algorithm based on multilayer information fusion | |
| Yang et al. | An effective and lightweight hybrid network for object detection in remote sensing images | |
| Ansari et al. | A survey of recent trends in two-stage object detection methods | |
| Tian et al. | Object detection in optical remote sensing images by integrating object-to-object relationships | |
| CN118247645A (en) | Novel DDCE-YOLOv s model underwater image target detection method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181130 |
|
| RJ01 | Rejection of invention patent application after publication |