Glass defect visual detection algorithm

Abstract

The invention relates to a glass defect visual detection algorithm. The algorithm comprises the following steps: aiming at different defects, carrying out different lighting so as to respectively obtain a plurality of channel images of each defect during different lighting; respectively selecting an image extraction edge system which is most suitable for extracting a channel image of the edge system for each defect; applying the selected image extraction edge system to the image of the channel so as to divide the channel image of each channel into a window area and an edge area; labeling the defects of the edge area and the window area; screening out small image blocks with defects by utilizing a machine vision algorithm and reserving the small image blocks for the deep learning model forfurther analysis; utilizing the defect detection model and the classification model to detect defect positions and attributes with categories; according to the defect attributes of the defects, training a grade discriminator through machine learning, and dividing the defect visibility into multiple grades from weak to strong. If defect categories need to be added, only annotation training is needed, a classifier does not need to be additionally designed, and product switching is conveniently achieved.

Description

technical field [0001] The invention relates to the technical field of glass defect detection, in particular to a glass defect visual detection algorithm. Background technique [0002] At present, the known mobile phone glass detection algorithms mainly rely on traditional machine vision algorithms, and use artificially constructed visual features to find glass defects that meet specific conditions. However, purely traditional detection algorithms can no longer meet the growing demand for glass detection: [0003] (1) Good versatility. When switching products, the algorithm needs to quickly adapt to new defect types. Due to different processes and imaging between different products, there will be various differences between different product images. Customized pure traditional vision algorithms have flexibility Poor, long adjustment period, unsatisfactory effect, etc.; [0004] (2) The detected defects need to be distinguished from strong to weak, because in actual product...

AI Technical Summary

Problems solved by technology

[0005] (3) Defects need to be classified into categories. In order to ensure product yield, some categories of defects can be let go, and some categories of defects should not be missed as much as possible. This technology can continuously train and learn new categories through deep learning to predict the type of defects. However, traditional algorithms tend to perform poorly when faced with a large number of defect categories.

Images