CN-119887664-B - Industrial topology type defect detection, classification and segmentation method

Abstract

The application discloses a method for detecting, classifying and dividing defects of an industrial topology type, which relates to the technical field of deep learning and digital image processing, and comprises a topological structure feature extraction module, a multi-view feature fusion module and a loss constraint and division module; the feature extraction module extracts multistage features from an input image to obtain topology feature information of defects, the multi-view feature fusion module supplements attention of features from multiple angles in a feature fusion process, retains important information from different global forms, and the loss constraint and segmentation module constrains loss of continuity based on continuous homology so as to better constrain segmented topology continuity and obtain a significant defect prediction graph. The topology type defect identification method and the topology type defect identification device can efficiently identify topology type defects.

Inventors

• Mai Yanheng
• ZHENG JIAQI
• ZHOU YANBO
• WANG XINJUE
• XIA YULONG

Assignees

• 苏州辰瓴信息技术有限公司

Dates

Publication Date

20260508

Application Date

20241226

Claims (3)

1. 1. The industrial topology type defect detection, classification and segmentation method is characterized by comprising a topological structure feature extraction module, a multi-view feature fusion module and a loss constraint and segmentation module, wherein the feature extraction module extracts multi-level features from an input image to obtain defective topological structure feature information, the multi-view feature fusion module supplements attention to the features from multiple angles in a feature fusion process, retains important information from different global forms, the loss constraint and segmentation module better constrains segmented topological continuity based on continuous homology constraint loss to obtain a significant defect prediction graph, the topological structure feature extraction module comprises a dynamic deformable convolution operator, the feature extraction module carries out two-dimensional change on the dynamic deformable convolution operator by introducing an offset to focus on tubular structures and curved local features, the multi-view feature fusion module comprises a decoupling head module and a topological continuity constraint loss module, and the decoupling head module can observe the structural features of the defects from multiple angles, and the topological continuity constraint loss module can realize the continuous constraint and continuous constraint loss of the topological constraint area by up-sampling and improves the spatial resolution, and the topological constraint loss constraint and the topological continuity constraint loss module has the effect of continuously limiting the topological constraint loss of a network; The method comprises the steps of firstly, carrying out two-dimensional change on a dynamic deformable convolution operator by introducing offset to focus on local characteristics of a tubular structure and bending, extracting multi-level characteristics from an input image by using the dynamic deformable convolution operator to obtain topology characteristic information of defects, and carrying out two-dimensional change specific operation on the dynamic deformable convolution operator in the first step, wherein for a given two-dimensional convolution coordinate, the center coordinate is M i = (u i , v i , one 3×3 convolution kernel M is represented as M= { (u-1, v-1), (u-1, v), (u+1, v+1) }, carrying out linearization on a standard convolution kernel in an x-axis and a y-axis, wherein in the u-axis direction, the specific position of each grid is M i±e = (u i±e , v i±e , in the v-axis direction, the specific position of each grid is M j±e = (u j±e , v j±e ), and e=0, 1,2,3,4 represents the horizontal distance from a center grid, wherein the deformation offset delta= { delta epsilon [ -1,1] } is M i+1 , and the offset is added up by adding Mi; The change in the u-axis direction is: the change in the v-axis direction is: Step two, the multi-view feature is fused with a plurality of morphological kernel templates generated based on dynamic deformable convolution, and structural features of defects are observed from multiple angles, wherein the multi-view feature fusion strategy for a given deformable convolution in the step two specifically operates as follows: Wherein g m (P u ) and g m (P v ) represent feature maps extracted from the u-axis and v-axis, respectively, of the deformable convolution operator; step three, the fusion result of each layer processed in the step two is transmitted into a decoupling head module to perform decoding, defect positioning and segmentation operation, and finally a significant defect prediction graph is obtained; And step four, calculating a loss value.
2. 2. The method for detecting, classifying and segmenting industrial topology type defects according to claim 1, wherein a random discard strategy is introduced in the second step to set a random discard probability p for each feature in each layer of feature map.
3. 3. The method of claim 2, wherein the topology continuity constraint loss in step four is the difference between the true value y and the predicted value y.

Description

Industrial topology type defect detection, classification and segmentation method Technical Field The application relates to the technical field of deep learning and digital image processing, in particular to a method for detecting, classifying and dividing defects of industrial topology types. Background With rapid advances in deep learning techniques and visual attention models, salient object detection, classification, segmentation have become a key area in computer vision research. The technology mimics the characteristics of the human visual system through a deep learning approach, aimed at efficiently identifying and locating specific objects in images or videos. The human visual system is able to selectively focus on regions of interest based on previous knowledge, which has inspired the development of saliency detection, classification, segmentation techniques. Significance detection, classification and segmentation are not only crucial for target recognition, but also show great application potential in various aspects of image and video compression, image retrieval, image redirection and the like. The infrastructure of modern neural networks dates back to 1998, when back propagation algorithms were introduced, enabling networks to learn optimization through back propagation errors while information is being passed forward. In recent years, deep learning models have been widely used in industrial defect detection, classification, and segmentation. For example, the advanced feature extraction neural network is used for detecting, classifying and dividing the defects of the metal surface, so that the detection, classification and division precision is effectively improved, the deep neural network has excellent performance in detecting, classifying and dividing the defects of the solar panel, and the real-time detection, classification and division model based on the real-time detection, classification and division frames is applied to rapid defect screening of packaging materials. The application of these techniques significantly improves the automation level and product quality of industrial production. In the manufacturing process of digital products such as mobile phones, the defects in or on the surfaces of tiny parts seem to be insignificant, but in practice, the performance and reliability of the products may be seriously affected. These defects may not only affect the electrical performance of the electronic component, resulting in unstable signal transmission, poor contact, etc., but may also become stress concentration points, reduce the mechanical strength of the assembly, and increase the risk of failure of the product during use. Therefore, it is important to detect, classify, and divide defects in such tiny parts. The efficient defect detection, classification and segmentation not only can help manufacturers to discover and correct problems in the production process in time and avoid defective products from flowing into markets, but also can remarkably improve the product quality and customer satisfaction, thereby enhancing the competitiveness of enterprises. Currently, the mainstream methods of defect detection, classification, segmentation remain manual visual inspection, and operators check defects on the surface of parts by visual inspection or by means of a magnifying glass or microscope. However, the efficiency of manual detection, classification and segmentation is low, and due to the personal quality and technical level difference of workers, the phenomena of missed detection and false detection are easy to occur, and the operation cost is high. Although the deep learning method is superior to the conventional method in terms of detection, classification, segmentation accuracy and efficiency, the application in the industrial field has not been popularized. Moreover, the accuracy and efficiency of detection, classification and segmentation are required to be improved, and the level of intellectualization and automation of industrial production is required to be improved. Disclosure of Invention The application aims to provide a method for detecting, classifying and dividing defects of industrial topology types, aiming at improving the technical problems. The technical scheme includes that the method comprises a topological structure feature extraction module, a multi-view feature fusion module and a loss constraint and segmentation module, wherein the feature extraction module extracts multi-level features from an input image to obtain topological structure feature information of the defect, the multi-view feature fusion module supplements attention of features from multiple angles in a feature fusion process, retains important information from different global forms, and the loss constraint and segmentation module constrains the segmented topological continuity better based on continuous homology constraint loss to obtain a significance defect prediction graph. Further, the topo