CN111126441A - Construction method of classification detection network model - Google Patents

Construction method of classification detection network model Download PDF

Info

Publication number
CN111126441A
CN111126441A CN201911167163.8A CN201911167163A CN111126441A CN 111126441 A CN111126441 A CN 111126441A CN 201911167163 A CN201911167163 A CN 201911167163A CN 111126441 A CN111126441 A CN 111126441A
Authority
CN
China
Prior art keywords
convolution
training
network model
convolution network
layer
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
CN201911167163.8A
Other languages
Chinese (zh)
Other versions
CN111126441B (en
Inventor
管声启
雷鸣
常江
倪弈棋
卢浩
郭飞飞
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.)
Xian Polytechnic University
Original Assignee
Xian Polytechnic 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 Xian Polytechnic University filed Critical Xian Polytechnic University
Priority to CN201911167163.8A priority Critical patent/CN111126441B/en
Publication of CN111126441A publication Critical patent/CN111126441A/en
Application granted granted Critical
Publication of CN111126441B publication Critical patent/CN111126441B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • 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

Abstract

The invention discloses a construction method of a classification detection network model, which comprises the following steps: acquiring a training sample of an image of an object to be classified, establishing a training model data set, inputting the training model data set into a first convolution network model for training to obtain a weight file; inputting the weight file into a second convolution network model to obtain a feature map and an original image corresponding to each layer of convolution network; inputting the feature map corresponding to each layer of the convolutional network and the original image into an image quality evaluation algorithm to obtain an evaluation result; selecting a proper convolution operation step length corresponding to each layer of the first convolution network model according to the evaluation result, and increasing the number of convolution kernels to form new convolution network parameters; and updating the first convolution network model by using the new convolution network parameters to obtain a classification network model. The requirements of flexible detection and intelligent detection are met, and the detection cost and the detection complexity are reduced.

Description

Construction method of classification detection network model
Technical Field
The invention belongs to the technical field of classification detection models, and relates to a construction method of a classification detection network model.
Background
With the development of big data and computer hardware, the neural network rises again, the technologies such as deep learning, artificial intelligence, big data, internet of things and the like begin to develop dramatically, and under the promotion of the development of computer technology, the manufacturing industry begins to gradually convert from the traditional mechanical production and the rigid manufacturing mode with heavy tasks into the automatic, intelligent and flexible intelligent manufacturing mode with a machine replacing the manual labor path. Artificial intelligence has become a field with many practical applications and active research topics, not just limited to the field of computers. Deep learning is the discipline of studying how computers simulate learning behaviors of human beings to acquire new knowledge or skills, and reorganize existing knowledge structures and continuously improve performance of the knowledge structures.
The concept of deep learning comes from the research of artificial neural networks, and a multilayer perceptron with multilayer hidden layers is a deep learning structure. Deep learning discovers a distributed feature representation of data by combining underlying information to form more abstract, high-level information to represent attribute categories or features. The motivation is to establish a mechanism for simulating the human brain to analyze and learn a human neural network and to interpret data such as images, texts, sounds and the like.
In the traditional mechanical field, computer vision or image processing technology is always adopted to solve the problem of rigid detection in the aspect of automatic image detection, but with the transformation of manufacturing industry, detection is taken as an important link of manufacturing, and the development of flexibility and intellectualization is also needed. However, the conventional detection technology is only a rigid detection link developed for a certain type of environment or a certain type of detection object. Therefore, in order to develop a flexible detection technology, a deep learning network with self-learning ability is used for the detection technology, and the flexible and intelligent development of a detection link is facilitated. At present, deep learning is rapidly developed in the mechanical industry, visual support is provided for industrial robots, and great progress is made in the aspects of image recognition, positioning, classification, measurement and the like. However, the fusion of deep learning and industrial technologies has not been able to truly realize flexible detection, and there is no classification network model that can be automatically adjusted for different detection objects.
Disclosure of Invention
The invention aims to provide a construction method of a classification detection network model, which can obtain the classification network model automatically adjusted according to different detection objects.
The invention adopts the technical scheme that the construction method of the classification detection network model comprises the following steps:
step 1, obtaining a training sample of an image of an object to be classified, establishing a training model data set, inputting the training model data set into a first convolution network model for training, and obtaining a weight file;
step 2, inputting the weight file into a second convolution network model to obtain a feature map and an original image corresponding to each layer of convolution network;
step 3, inputting the feature images corresponding to each layer of the convolutional network and the original image into an image quality evaluation algorithm to obtain an evaluation result;
step 4, selecting a proper convolution operation step length corresponding to each layer of the first convolution network model according to the evaluation result, and increasing the number of convolution kernels to form new convolution network parameters;
and 5, updating the first convolution network model by using the new convolution network parameters to obtain a classification network model.
The invention is also characterized in that:
the step 1 specifically comprises the following steps:
step 1.1, obtaining a training sample of an image of an object to be classified, establishing a training model data set, wherein the training model data set comprises a training set, a verification set and a prediction set, and generating a training set marking file, a verification set marking file and a prediction set marking file;
and step 1.2, inputting the marked files of the training set as training data and the marked files of the verification set as verification data into a first convolution network model for training to obtain a weight file.
The second convolutional network is a deconvolution network.
The step 2 specifically comprises the following steps: and the second convolution network model extracts the characteristic values stored in the convolution kernel filter corresponding to each layer of convolution network in the weight model file, projects the characteristic values to the corresponding pixel space to obtain the characteristic diagram of each layer of convolution network, extracts the output image corresponding to each layer of convolution network, and outputs the original image corresponding to the characteristic diagram.
The step 3 specifically comprises the following steps: and comparing the characteristic diagram serving as a distorted image with the original image to obtain a quality evaluation result of the characteristic diagram.
The invention has the beneficial effects that:
the method for constructing the classification detection network model comprises the steps of training a training model data set through a first convolution network model, deconvolving a second convolution network to obtain a feature map and an original image, and optimizing parameters of the classification network model according to the quality evaluation results of the feature map and the original image to obtain a new classification detection network model, wherein the method is suitable for constructing classification detection network models of different classification objects; the requirements of flexible detection and intelligent detection are met, and the detection cost and the detection complexity are reduced.
Drawings
FIG. 1 is a flow chart of a method of constructing a classification detection network model according to the present invention;
FIG. 2 is a schematic diagram of a VGG19 network model and a deconvolution network model in the construction method of the classification detection network model of the present invention;
fig. 3 is a diagram of evaluation results of a feature diagram and an original image in the method for constructing a classification detection network model according to the present invention.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
The invention relates to a construction method of a classification detection network model, which specifically comprises the following steps as shown in figure 1:
step 1, obtaining a training sample of an image of an object to be classified, establishing a training model data set, inputting the training model data set into a first convolution network model for training, and obtaining a weight file;
step 1.1, obtaining a training sample of an image of an object to be classified, establishing a training model data set, wherein the training model data set comprises a training set, a verification set and a prediction set, and generating a training set marking file, a verification set marking file and a prediction set marking file;
specifically, in this embodiment, nine hot rolled steels with surface defects are used as training samples, which are cr, gg, in, pa, ps, rp, rs, sc, and sp, and each of the hot rolled steels contains 200 pictures, and the size of each picture is 200 × 200 pixels, for a total of 1800 pictures. The data sets were divided into:
training set-train: nine defects are 150 in each, and 1350 in total;
validation set-val: nine defects were 25 sheets each, for a total of 225 sheets;
prediction set-predict: nine defects were 25 sheets each, for a total of 225 sheets.
The method comprises the following steps that nine types of defect training sets, a verification set and a prediction set all contain 9 files, nine types of defects are respectively marked as 0, 1, 2, 3, 4, 5, 6, 7 and 8 according to categories, and corresponding train. Then, the trail.label file and the trail.training set are input into the create.record.py function to obtain the trail _ tf.record, and the val.label and the val verification set generate a corresponding val _ tf.record file by using the create.record.py function in the same way.
And step 1.2, inputting the marked files of the training set as training data and the marked files of the verification set as verification data into a first convolution network model for training to obtain a weight file.
The first convolution network model is a VGG19 network model, a schematic diagram of the VGG19 network model is shown in fig. 2, and the VGG19 model is composed of sixteen convolution layers (conv1_1, conv1_2, conv2_1, conv2_2, conv3_1, conv3_2, conv3_3, conv3_4, conv4_1, conv4_2, conv4_3, conv4_4, conv5_1, conv5_2, conv5_3, conv5_4), four maximum pooling layers (Maxpooling), an Average pooling (Average pooling), a Full connection layer (Full connection) and a softconnection layer.
Specifically, two tf.record files are used as train training data input and val verification data input of a VGG19 classification model for iterative training, a weight file, namely a checkpoint file, of the VGG19 network for classifying the defects of the strip steel is obtained after training (the checkpoint file is in a model parameter file storage format obtained after model training based on a tensoflow frame, and the checkpoint file in the tensoflow is a binary file for storing values of all weights, biass, gradients and other variables), and then the weight file is converted into a npy file from the checkpoint format.
Step 2, inputting the weight file into a second convolution network model to obtain a feature map and an original image corresponding to each layer of convolution network;
the second convolution network model is a deconvolution network, the network structure is basically consistent with VGG19, the schematic diagram of the deconvolution network model is shown in FIG. 2, convolution and pooling operations adopt completely opposite operations, namely deconvolution (deconv) and deconvolution (pooling) operations, feature values stored in convolution kernel filters corresponding to each layer of convolution networks of the 16 layers of convolution networks in the weight model file obtained after training in step 1 are extracted and projected to corresponding pixel spaces to obtain feature diagrams corresponding to the 16 layers of convolution networks, and the deconvolution network can also extract output images image corresponding to each layer of convolution networks and output 16 original images (images) in one-to-one correspondence with the 16 feature diagrams.
Step 3, inputting the feature images corresponding to each layer of the convolution network and the original image into an image quality evaluation algorithm to obtain an evaluation result;
specifically, 16 feature maps (feature maps) corresponding to the 16-layer convolutional network are input into an image quality evaluation algorithm (SSIM) as a distorted image and 16 original images to obtain 16 evaluation results. These 16 corresponding results are also the quality assessment results of the 16-layer convolutional network signature and output maps. The quality of the feature map extracted by each layer of the convolutional network of the visual network is evaluated through a structural similarity image quality evaluation algorithm, so that the performance of extracting features of each layer of the convolutional network under the industrial background of the surface defects of the strip steel is judged. The worse the corresponding evaluation result is, the larger the convolution kernel selected by the filter of the corresponding convolution network is, the larger the convolution step length is, and the more the characteristic information loss is caused. In the convolution operation, the size of the convolution kernel determines the size of the image acquisition visual field, and the smaller the convolution kernel, the smaller the image acquisition visual field, the more data is acquired, and the larger the calculation amount is. The convolution step determines the translation length of the image acquisition visual field, and the larger the translation length is, the larger the span of the image acquisition visual field is, and the less information is acquired.
The structural similarity is an index for measuring the similarity of two images, and the structural similarity range is 0-1. When the two images are identical, the value of SSIM is equal to 1.
The SSIM structural similarity index defines information of an image structure as being independent of brightness (luminance), contrast (contrast), and attributes reflecting an object structure (structure) in a scene from the viewpoint of image composition, and models distortion as a combination of three different factors of brightness contrast (luminance contrast), contrast (contrast), and structure contrast (structure contrast). The mean value is used as the luminance estimate, the standard deviation is used as the contrast estimate, and the covariance is used as the measure of the degree of structural similarity.
In fig. 3, the comparison result of the evaluation image of the feature map (feature map) and the output map (image) extracted from the 16-layer convolutional network obtained by training the strip steel defect data set by using two image quality evaluation algorithm peak signal-to-noise ratio algorithm PSNR and structural similarity image quality evaluation algorithm SSIM is shown in the figure, and finally, the SSIM image evaluation quality algorithm is used for the image quality evaluation module of the present invention.
Step 4, selecting a proper convolution operation step length corresponding to each layer of the first convolution network model according to the evaluation result, and increasing the number of convolution kernels to form new convolution network parameters;
specifically, two classification decisions are performed on 16 evaluation results output by the image evaluation module one by one, a proper convolution operation step length corresponding to each layer of convolution network is selected, the number of convolution kernels is increased, and new convolution network parameters are formed in an instructive manner. The reason for the poor image quality evaluation result is that the loss of the characteristic information is large, and the corresponding parameter setting of the convolution network is unreasonable, for example, the loss of the image pixel points is caused by adopting a large convolution kernel and a convolution step length, or the number of the corresponding convolution kernels is small, so that the image quality is reduced. Therefore, a new network model suitable for the current classification data set is generated according to the result of the image quality evaluation and used as an index for classifying the optimization of the network model parameters.
And 5, updating the first convolution network model by using the new convolution network parameters to obtain a classification network model. The obtained classification network model can be directly used for classifying the defects of the strip steel, and the classification network models of different classification detection objects can be obtained only by changing the training sample to obtain the training data set.
According to the method for constructing the classification detection network model, the training model data set is trained through the first convolution network model, the second convolution network deconvolves to obtain the feature map and the original image, and the classification network model parameter optimization is carried out according to the feature map and the original image quality evaluation result to obtain a new classification detection network model, so that the method is suitable for constructing the classification detection network models of different classification objects; the requirements of flexible detection and intelligent detection are met, and the detection cost and the detection complexity are reduced.

Claims (5)

1. A construction method of a classification detection network model is characterized by comprising the following steps:
step 1, obtaining a training sample of an image of an object to be classified, establishing a training model data set, inputting the training model data set into a first convolution network model for training, and obtaining a weight file;
step 2, inputting the weight file into a second convolution network model to obtain a feature map and an original image corresponding to each layer of convolution network;
step 3, inputting the feature images corresponding to each layer of the convolution network and the original image into an image quality evaluation algorithm to obtain an evaluation result;
step 4, selecting a proper convolution operation step length corresponding to each layer of the first convolution network model according to the evaluation result, and increasing the number of convolution kernels to form new convolution network parameters;
and 5, updating the first convolution network model by using the new convolution network parameters to obtain a classification network model.
2. The method for constructing a classification detection network model according to claim 1, wherein step 1 specifically includes:
step 1.1, obtaining a training sample of an image of an object to be classified, establishing a training model data set, wherein the training model data set comprises a training set, a verification set and a prediction set, and generating a training set marking file, a verification set marking file and a prediction set marking file;
and step 1.2, inputting the marked files of the training set as training data and the marked files of the verification set as verification data into a first convolution network model for training to obtain a weight file.
3. The method of claim 1, wherein the second convolutional network is a deconvolution network.
4. The method for constructing a classification detection network model according to claim 1, wherein the step 2 specifically includes: and the second convolution network model extracts the characteristic values stored in the convolution kernel filter corresponding to each layer of convolution network in the weight model file, projects the characteristic values to the corresponding pixel space to obtain the characteristic diagram of each layer of convolution network, extracts the output image corresponding to each layer of convolution network, and outputs the original image corresponding to the characteristic diagram.
5. The method for constructing a classification detection network model according to claim 1, wherein the step 3 specifically includes: and comparing the characteristic diagram serving as a distorted image with the original image to obtain a quality evaluation result of the characteristic diagram.
CN201911167163.8A 2019-11-25 2019-11-25 Construction method of classification detection network model Active CN111126441B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201911167163.8A CN111126441B (en) 2019-11-25 2019-11-25 Construction method of classification detection network model

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201911167163.8A CN111126441B (en) 2019-11-25 2019-11-25 Construction method of classification detection network model

Publications (2)

Publication Number Publication Date
CN111126441A true CN111126441A (en) 2020-05-08
CN111126441B CN111126441B (en) 2023-04-07

Family

ID=70496579

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201911167163.8A Active CN111126441B (en) 2019-11-25 2019-11-25 Construction method of classification detection network model

Country Status (1)

Country Link
CN (1) CN111126441B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111860582A (en) * 2020-06-11 2020-10-30 北京市威富安防科技有限公司 Image classification model construction method and device, computer equipment and storage medium

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108491880A (en) * 2018-03-23 2018-09-04 西安电子科技大学 Object classification based on neural network and position and orientation estimation method
CN108830242A (en) * 2018-06-22 2018-11-16 北京航空航天大学 SAR image targets in ocean classification and Detection method based on convolutional neural networks
US20190102646A1 (en) * 2017-10-02 2019-04-04 Xnor.ai Inc. Image based object detection

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190102646A1 (en) * 2017-10-02 2019-04-04 Xnor.ai Inc. Image based object detection
CN108491880A (en) * 2018-03-23 2018-09-04 西安电子科技大学 Object classification based on neural network and position and orientation estimation method
CN108830242A (en) * 2018-06-22 2018-11-16 北京航空航天大学 SAR image targets in ocean classification and Detection method based on convolutional neural networks

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
杨洁等: "基于卷积网络的视频目标检测", 《南华大学学报(自然科学版)》 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111860582A (en) * 2020-06-11 2020-10-30 北京市威富安防科技有限公司 Image classification model construction method and device, computer equipment and storage medium

Also Published As

Publication number Publication date
CN111126441B (en) 2023-04-07

Similar Documents

Publication Publication Date Title
CN108564029B (en) Face attribute recognition method based on cascade multitask learning deep neural network
CA3043621C (en) Method and system for color representation generation
CN109685072B (en) Composite degraded image high-quality reconstruction method based on generation countermeasure network
CN113240580A (en) Lightweight image super-resolution reconstruction method based on multi-dimensional knowledge distillation
CN110738697A (en) Monocular depth estimation method based on deep learning
CN106920243A (en) The ceramic material part method for sequence image segmentation of improved full convolutional neural networks
CN111310862A (en) Deep neural network license plate positioning method based on image enhancement in complex environment
CN110647875B (en) Method for segmenting and identifying model structure of blood cells and blood cell identification method
CN107423747A (en) A kind of conspicuousness object detection method based on depth convolutional network
CN108985342A (en) A kind of uneven classification method based on depth enhancing study
CN105243385B (en) A kind of image quality evaluating method based on unsupervised learning
CN112819063B (en) Image identification method based on improved Focal loss function
Zhang et al. Inkthetics: a comprehensive computational model for aesthetic evaluation of Chinese ink paintings
CN109543749A (en) Drawing sentiment analysis method based on deep learning
CN115222754A (en) Mirror image segmentation method based on knowledge distillation and antagonistic learning
CN116543227A (en) Remote sensing image scene classification method based on graph convolution network
CN111126441B (en) Construction method of classification detection network model
CN115240259A (en) Face detection method and face detection system based on YOLO deep network in classroom environment
CN113361496B (en) City built-up area statistical method based on U-Net
Pen et al. Boundary precedence image inpainting method based on self-organizing maps
CN114330234A (en) Layout structure analysis method and device, electronic equipment and storage medium
CN113688789B (en) Online learning input degree identification method and system based on deep learning
CN114972282A (en) Incremental learning non-reference image quality evaluation method based on image semantic information
CN111369124A (en) Image aesthetic prediction method based on self-generation global features and attention
CN112115779A (en) Interpretable classroom student emotion analysis method, system, device and medium

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