CN110889461A

CN110889461A - Image multi-feature extraction and classification method for chassis assembly quality detection

Info

Publication number: CN110889461A
Application number: CN201911248305.3A
Authority: CN
Inventors: 刘桂雄; 何彬媛
Original assignee: South China University of Technology SCUT
Current assignee: South China University of Technology SCUT
Priority date: 2019-12-09
Filing date: 2019-12-09
Publication date: 2020-03-17
Anticipated expiration: 2039-12-09
Also published as: CN110889461B

Abstract

The invention provides an image multi-feature extraction and classification method for chassis assembly quality detection, which comprises the following steps: introducing sparse clustering, and constructing a case assembly detection classification model; adopting large-size convolution decomposition, enlarging the scope of receptive field and reducing the number of parameters at the same time, and obtaining the multiple characteristics of the case; based on the Incepton structure, a four-branch multi-scale convolution kernel is adopted, so that the calculation performance is improved; and (4) multi-feature connection, inputting the classification into a classifier and finally realizing classification of the type of the case and the parts. The invention utilizes the Incep structure feature extraction and characterization advantages, solves the problems of information loss and the like of complex image feature extraction by a classical image classification method, enlarges the receptive field range by setting a multi-branch structure, reduces the parameter quantity by adopting rolling and decomposition, and is beneficial to application in visual detection of chassis assembly quality identification.

Description

Image multi-feature extraction and classification method for chassis assembly quality detection

Technical Field

The invention relates to the field of assembly quality detection, in particular to an image multi-feature extraction and classification method for chassis assembly quality detection.

Background

The visual inspection technology is widely applied because the production efficiency can be improved, the robot is replaced by a robot, and the upgrading of the assembly industry is promoted. As the types of the chassis are various, the chassis assembly standard part has the characteristics of multiple parts, complex types and high similarity of the same type of parts, and the manual inspection mode is difficult to meet the requirement of mass production. The quality of the assembly quality of the case directly influences the use of products, and the quality of the case is very necessary to be detected and analyzed. The features adopted based on the classic image classification algorithm are all image bottom layer visual features, the pertinence to specific images and specific classification modes is not sufficient, the classification precision is greatly reduced for the problems of slight difference between categories, serious image interference and the like, and the classic image classification method cannot achieve a good effect in a complex scene.

In recent years, with the development of visual inspection technology, the huge improvement of computing power and the continuous expansion of deep learning methods, a deep network is applied and developed on a visual inspection task, and a convolutional neural network is applied to various industries, particularly the quality inspection of manufactured products, and achieves good effect. The convolutional neural network extracts image features layer by layer, realizes image classification through the classifier according to the multi-feature statistical data, finishes a classification detection task, and has the characteristics of various extracted features, high classification accuracy and the like. It can be seen that the traditional visual detection method for the quality of manufactured products has poor effect and low efficiency, and meanwhile, the technology for detecting the quality of manufactured products based on deep learning is a future trend in the field. If the multi-feature extraction and characterization advantages of the convolutional neural network can be utilized, the problem that the classical image classification algorithm is insufficient in pertinence to complex images and specific classification modes is solved, the large-size convolution and Incep structure is utilized, the receptive field is enlarged, the calculated amount is reduced, and the application in visual detection and classification of the assembly quality of the chassis is facilitated.

Disclosure of Invention

In order to solve the problems and the defects, the invention provides an image multi-feature extraction and classification method for chassis assembly quality detection, which utilizes the multi-feature extraction and characterization advantages of a convolutional neural network to solve the problem of insufficient pertinence of a classical image classification algorithm to a complex image and a specific classification mode, utilizes a large-size convolution and Incep structure to enlarge the receptive field and reduce the calculated amount, and is beneficial to application in chassis assembly quality visual detection and classification.

The purpose of the invention is realized by the following technical scheme:

a chassis assembly quality detection-oriented image multi-feature extraction and classification method comprises the following steps:

introducing sparse clustering, and constructing a case assembly detection classification model;

b, adopting large-size convolution decomposition, enlarging the scope of a receptive field and reducing the number of parameters at the same time, and obtaining multiple characteristics of the case;

c, based on an Incepration structure, adopting a four-branch multi-scale convolution kernel to improve the calculation performance;

d, multi-feature connection, input classifier and final classification of case types and parts.

The invention has the beneficial effects that:

the problem that the conventional image classification algorithm is insufficient in pertinence to complex images and specific classification modes is solved by utilizing the multi-feature extraction and characterization advantages of the convolutional neural network, the receptive field is enlarged and the calculated amount is reduced by utilizing large-size convolution and an Incep structure, and the method is favorable for being applied to visual detection and classification of the assembly quality of the chassis.

Drawings

FIG. 1 is a flow chart of an image multi-feature extraction and classification method for chassis assembly quality detection.

Detailed Description

The present invention will be described in further detail with reference to the following examples and accompanying drawings.

The invention relates to an image multi-feature extraction and classification method for chassis assembly quality detection, which comprises the following steps of:

step 10, introducing sparse clustering, and constructing a case assembly detection classification model:

case assembly detection classification model is optimized by introducing sparse clustering

Wherein S, β, y and lambda respectively represent a feature array, a sparse vector, a sparse observation and a control coefficient.

Step 20, adopting large-size convolution decomposition to enlarge the receptive field range and reduce the number of parameters at the same time, and obtaining the multiple characteristics of the case:

and (3) adopting large-size convolution decomposition, and replacing 3 × 3 convolution with 3 × 1 convolution to obtain a triple Loss function:

in the formula, a and p are in the same class, a and n are in different classes, and + represents that the internal value of [ ] is greater than zero, the value is taken as the loss value, and the loss is zero when the internal value is less than zero.

For the parameter quantity increase caused by the large convolution expansion receptive field, large convolution decomposition is adopted, and the gradient formula for reducing the calculated quantity is as follows:

and (3) decomposing the symmetric convolution kernel into an asymmetric convolution kernel by adopting convolution decomposition, so that the calculated amount is reduced:

where n is the convolution kernel size.

Step 30, based on the inclusion structure, a four-branch multi-scale convolution kernel is adopted to improve the calculation performance:

based on the Incep structure, a full connection layer is changed into sparse connection, convolution kernels with different sizes are adopted in four branches, sparse structures are constructed by 1 × 1, 1 × 3, 3 × 1, 3 × 3 and the like, different convolution kernels are adopted to obtain different sizes of receptive fields, and the chassis features with different sizes are obtained. The convolutional neural network of the embodiment is connected by 20 layers of inclusion modules: comprises 1 × 1, 1 × 3, 3 × 1, 3 × 3 convolution layer sparse connection; 1 × 1, 3 × 3, 5 × 5 convolutional layer sparse connections.

Step 40, multi-feature connection, input classifier and final case type and part classification:

and C, extracting the high-level and low-level multi-features from the case assembly picture in the steps B and C, connecting the features, and inputting the features into a classifier to finally realize the classification of the case types and the parts. In the embodiment, the inclusion modules are sparsely connected and then input into the 7 x 7 average pooling layer and then input into the full-connection layer, and the result output by the classifier is the case assembly quality detection classification result.

The method utilizes the multi-feature extraction and characterization advantages of the convolutional neural network, solves the problem that the classical image classification algorithm is insufficient in pertinence to complex images and specific classification modes, utilizes large-size convolution and an inclusion structure, enlarges the receptive field, reduces the calculated amount, and is beneficial to application in visual detection and classification of the assembly quality of the case.

Although the embodiments of the present invention have been described above, the above descriptions are only for the convenience of understanding the present invention, and are not intended to limit the present invention. It will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the invention as defined by the appended claims.

Claims

1. An image multi-feature extraction and classification method for chassis assembly quality detection is characterized by comprising the following steps:

2. The image multi-feature extraction and classification method for chassis assembly quality detection according to claim 1, wherein in the step A, a chassis assembly detection classification model is optimized by introducing sparse clustering

Wherein S, β, y and lambda represent the feature array, sparse vector, sparse observation and control system respectivelyAnd (4) counting.

3. The chassis assembly quality inspection-oriented image multi-feature extraction and classification method of claim 1, wherein in the step B, a large-size convolution decomposition is adopted, and a 3 x 1 convolution is used to replace a 3 x 3 convolution, so as to obtain a triple Loss function:

4. The image multi-feature extraction and classification method for chassis assembly quality detection according to claim 1, wherein in the step B, for the parameter quantity increase caused by the large convolution expansion receptive field, convolution decomposition is adopted, and a gradient formula for reducing the calculated quantity is as follows:

5. the chassis-assembly-quality-detection-oriented image multi-feature extraction and classification method of claim 1, wherein in the step B, a large convolution decomposition is adopted to decompose a symmetric convolution kernel into an asymmetric convolution kernel, so that the calculation amount is reduced:

where n is the convolution kernel size.

6. The image multi-feature extraction and classification method for chassis assembly quality detection according to claim 1, wherein in the step C, based on an inclusion structure, a full connection layer is changed into sparse connection, convolution kernels of different sizes are adopted for four branches, sparse structures are constructed for 1 × 1, 1 × 3, 3 × 1 and 3 × 3, different reception fields are obtained by adopting different convolution kernels, and chassis features of different sizes are obtained.

7. The image multi-feature extraction and classification method for chassis assembly quality detection according to claim 1, wherein in the step D, the chassis assembly picture is subjected to high-level and low-level multi-features extraction in the steps B and C, and finally the features are connected and input into a classifier to finally realize the classification of the chassis type and the parts.