Use method of neural network classifier for machine vision defect detection

Abstract

The invention discloses a use method of a neural network classifier for machine vision defect detection, and belongs to the technical field of machine vision defect detection. The operation process of the method comprises classifier training and classifier reasoning. The classifier training process comprises the steps of processing images of defect samples, improving an original classification mode, enabling a plurality of ROI images to participate in training together, and forming a more accurate and efficient classifier. The classifier reasoning process is used for clearly detecting the station with the abnormal condition, the rule of the classifier is conveniently understood and modified, the method is scientific and reasonable, the traditional image processing technology is combined, a mask mode containing multiple ROIs is used, single-label classification of the image is converted into multi-label classification, and the method is suitable for large-scale popularization and application. Therefore, the problem that classifier discrimination vulnerabilities are difficult to find is obviously improved, project implementation is promoted, and customers are helped to quickly meet the requirement of defect detection.

Description

technical field [0001] The invention relates to the technical field of machine vision defect detection, in particular to a method for using a neural network classifier for machine vision defect detection. Background technique [0002] In the field of machine vision defect detection, the application of deep learning is an important means to solve the problem of product detection with complex appearance, and deep neural network classifier is the most commonly used and basic type of deep learning application, but in industrial defect detection , deep neural network classifiers also have many deficiencies. [0003] First of all, there are multiple detection areas in the product, and the discrimination standards of each area are different, so there are many defect combinations, and it is difficult to collect comprehensive samples in the actual production process. Due to the lack of sufficient and comprehensive training samples, the classifier often cannot If it is fully learned,...

AI Technical Summary

Problems solved by technology

[0003] First of all, there are multiple detection areas in the product, and the discrimination standards of each area are different, so there are many defect combinations, and it is difficult to collect comprehensive samples in the actual production process. Due to the lack of sufficient and comprehensive training samples, the classifier often cannot is fully learned, it is prone to misjudgment; secondly, it is difficult for a trained classifier to adjust the judgment rules as quickly as a human being; and it is difficult for humans to intuitively evaluate the classifier's understanding of the rules

[0004] When a complex classifier gives the category judgment standard, it is often difficult for humans to know which part of the information in the image the classifier is based on. Although the target detector can be used to detect the target, doing so will make the classification problem Transformed into an object detection problem

Images