Highly reflective surface defect detection method based on image processing and neural network classification

Abstract

The invention discloses a high-reflection surface defect detection method based on image processing and neural network classification. The original collected image is first subjected to basic processing such as background subtraction and denoising, and the parts that may have defects are preliminarily determined through feature extraction. A series of regional images of corresponding parts are obtained, and then the image sequence is input into the neural network classifier obtained by using the local feature area tiles as training data to judge whether it is a true defect, and the output result of the classifier is used as the final judgment result. The invention proposes a surface defect detection method of a highly reflective test piece, which uses the front-end two-dimensional digital image processing module to search and extract defect features, and combines the back-end neural network classifier for feature filtering to enhance the accuracy of search results. Under the premise of fully extracting the surface defects of the test piece, the probability of wrong and missed detection in the traditional image processing detection method can be reduced, and at the same time, the computing efficiency and versatility can be taken into account.

Description

technical field [0001] The invention relates to pattern classification technology, in particular to machine vision two-dimensional image processing technology and pattern classification technology based on neural network deep learning. Background technique [0002] In the production process of parts with highly reflective surfaces such as metal, bumps and scratches on the surface of parts are inevitable in the production line environment, which creates a need for rapid and effective identification and classification of defective and qualified products. The non-contact detection method based on machine vision, with the image processing theory as the core, has the advantages of high efficiency, low probability of false detection and avoiding secondary damage to the tested part, and has been widely used in the field of non-destructive testing. [0003] One of the common machine vision intelligent non-destructive testing methods is to directly perform feature extraction and anal...

AI Technical Summary

Problems solved by technology

[0003] One of the common machine vision intelligent non-destructive testing methods is to directly perform feature extraction and analysis based on digital image processing algorithms on the collected images of the tested part, and then make a qualified or unqualified judgment, which has the advantages of easy implementation and simple model. However, this method has high requirements on the quality of the collected images (sharpness, contrast, and brightness) and the environment of the collection site. It is necessary to manually build and debug the parameters of the detection model algorithm according to the specific application environment to ensure high robustness. However, this also limits the application scenarios of the model; there is also a surface detection method based on laser three-dimensional scanning, which uses the surface point cloud data obtained by laser scanning to analyze the characteristics of the surface of the tested part, but this method is in When the surface features of the tested part are subtle, the requirements for point cloud scanning accuracy and processing algorithm are very high, and the data processing is difficult and difficult to implement; the classifier detection method based on machine learning, such as Support Vector Machine SVM (Support Vector Machine ), convolutional neural network CNN (Convolutional Neural Network), etc., are also used in the field of non-destructive testing. Generally, the image data set of the tested surface is directly used as the training data, but this method is more suitable for the defect feature is very obvious or the tested In the case of less surface topography information (no too many details, such as grooves, bumps, lines, etc.), when the shape of the measured surface is complex or the target features are too subtle compared with the shape of the measured piece, the neural network classification The training of the machine is usually more difficult and requires a longer training and adjustment cycle

Images