CELL-based edge defect detection method

Abstract

The invention relates to a CELL-based edge defect detection method, which comprises the following steps: collecting accumulated data samples of historical detection CELL products to obtain data labeling samples; inputting the data annotation sample into a deep convolutional neural network for model training, and training a stably convergent segmentation model; capturing a product image of a CELL product to be detected, segmenting a defect coordinate position and a mark coordinate position in the product image through the segmentation model, and outputting the defect coordinate position; calling a product image of the CELL product to be detected, and performing linear fitting on the limited number of edge points to obtain a horizontal fitting line and a vertical fitting line respectively; and according to the mark coordinate position, carrying out spacing measurement on the fitted horizontal fitting line and vertical fitting line to obtain spacing parameter output. According to the invention, the segmentation model is established, the defects in the input product image can be segmented, the defects and the edge distance of the CELL product can be measured synchronously, the detection efficiency is improved, and the defect and edge distance detection precision is high.

Description

Technical field [0001] The present invention relates to the technical field of detection CELL edge defects, particularly to a an edge-based method for detecting defects CELL. Background technique [0002] Industrial machine vision is very common quality testing of a technology, mainly for industrial products photographed by the camera, then analyzed for product images by image processing technology, which by observing the real imaging products in the production process, the product whether or not there there are flaws and problems. CELL The present invention describes a method based on the traditional edge detection techniques and learning algorithms depth of the mixing process, to make up for poor robustness traditional algorithms, parameters more difficulties. [0003] A long period of time, means for machine vision applications are using conventional image processing techniques for quantitative image analysis, it is determined whether the defect exists. Since the industrial pr...

AI Technical Summary

Problems solved by technology

The biggest problem with pattern recognition technology is that it requires humans to manually extract and summarize features, but humans can only summarize those abstract features that are similar to human perception, and can only extract those indicators that are very easy to quantify

For the problem of defect classification, the features are often the length and width of the defect, gray value, contrast, duty cycle, area, etc. These features can only classify and judge the defects of conventional normal shapes, and it is difficult to have very strong robustness. Stickiness, for example, the human eye can distinguish two aspect ratios greatly, but defects belonging to the same category can easily be classified into one category, but if handed over to a machine, it may be judged as two different types based on the summarized features category

Images