A product surface defect detection method, system, device and medium

Abstract

The invention discloses a product surface defect detection method, system, device and medium, comprising: constructing a picture reconstruction network; collecting training pictures for a defect-free picture method; using the training pictures to train the picture reconstruction network to obtain the trained picture reconstruction network; collect pictures of surface defects of products to be processed, input the pictures of surface defects of products to be processed into the image reconstruction network after training, calculate the eigenvalues ​​of pictures of surface defects of products to be processed and reconstructed pictures, and calculate surface defects of products to be processed The difference between the eigenvalues ​​of the picture and the reconstructed picture, the product surface defect detection result is obtained based on the difference of the eigenvalues; the present invention only uses the good picture for surface defect detection, and the present invention is based on the full convolutional neural network, by reconstructing the good picture. Learn the ability to repair surface defects, so as to achieve pixel-level recognition capabilities for various surface defects.

Description

technical field [0001] The invention relates to the technical field of intelligent manufacturing and artificial intelligence, and in particular, to a method, system, device and medium for detecting surface defects of products. Background technique [0002] Surface defect detection algorithms based on machine vision mainly include traditional vision methods and deep learning methods. Traditional visual methods are mainly based on artificially designed visual features, such as watershed algorithm, LBP feature method, etc., while deep learning methods are mainly based on convolutional neural networks, such as Faster RCNN algorithm, MASK RCNN algorithm, etc. The use of deep learning methods in automatic surface defect detection algorithms is currently gaining popularity. [0003] In the current surface defect detection methods using deep learning, supervised learning methods such as target detection algorithms or semantic segmentation algorithms are mainly used. This type of m...

AI Technical Summary

Problems solved by technology

This type of method mainly faces four problems in practical application. First, it needs to collect a large number of defect pictures as training data. The collection of low-probability defect pictures in actual production requires high time and labor costs; second, based on the existing The automatic surface defect detection algorithm for training pictures cannot cope with the unknown defect manifestations that may occur in production. The method based on supervised learning can achieve a better generalization effect on the defect forms contained in the training data set, while the actual production The sudden unknown new defects in the process do not exist in the existing training pictures, and the algorithm may miss detection of defects, resulting in production losses; third, a large number of manually labeled defect pictures, object detection and semantic segmentation, etc. The cost is high and the time required is long, which leads to the slow response of the algorithm to changes in production conditions and poor timeliness; fourth, it is difficult to adapt to new product pictures. In supervised learning, good product pictures without defects are used as negative For learning from pictures, the background of new product pictures that do not exist in the training data may change greatly, which may cause the algorithm to over-check the defects. To solve this problem, it is necessary to invest in manpower to collect and label new product pictures in time , and then optimize the model, resulting in slower algorithm update

Images