Research And System Design Of Workpiece Defect Detection Method Based On CNN

The products usually contain some flaws on its surface in the process of product manufacturing, due to various factors affecting production materials and manufacturing equipments, with which the quality safety and operational performance of the manufactured workpieces will be destroyed. The traditional manual detection method for these flaws has been unable to catch up with speed of production, and has not ensure the quality of detecting work on account of subjective factors and mental state. In recent years, with the development of machine vision technology, workpiece defect detection systems by machines are gradually put into use. The workpiece defect detection systems have promoted the development of industrial automation and intelligence, but a good robustness one is still a challenging problem. Therefore, the study of new workpiece defect detection methods and systems would have important meanings both in theory and applications.In recent years, the convolutional neural network (CNN) exerts potential values in field of the industrial detection, medical analysis and traffic monitoring, etc. While most of the existing techniques have focused on building a robust network model, requiring a process of the feature representation and classifier building, the CNN model includes feature extraction stages in its own structures, which make the extracted features fully reflect the characteristics of the image information, and save the work of selecting features combination. Moreover, the manual extraction characteristics cannot be simply adjusted according to the new data set, but by using its own learning ability CNN can get trainable features and adapt to these changes naturally.This paper firstly outlines the development process of workpiece defect detection systems, and compares the relative techniques. On this basis, the workpiece defect detection system based on CNN is proposed to solve the problem of small defects detection. Aiming at the difficults of getting the workpiece mark relative position and the low rate of defect detection, this paper puts forward "two-times identification and two-times detection" defect detection scheme. Two-times identification is implemented by template matching method and CNN model, and two-times detection is used to enhance accuracy by comparing the identification results with the re-inspection. Secondly, a CNN defect detection method is put forward by analyzing the data of the defected and the non-defected images in the output node of CNN. The concept of resolution ratio is then defined by the similarity relationship between these data, which is used to detect the fine defects in the workpiece which cannot be checked out in the last step. Finally, the workpiece defect detection system is implemented, which achieves the real-time detection of the defected workpieces, and have the ability to show the information of defected images and defected locations, etc. Though extensive experiment results, the detection precision and detection speed are proved to demonstrate the effectivity and feasibility of the developed workpiece defect detection system based on CNN, in which the defect detection rate reaches 93.3%. In the end of the thesis, the error reasons are analyzed fully and the future study direction is also discussed.