Research On Surface Defects Detection System Of Reflected Surface Based On Deep Learning

Machine vision based surface defect detection technology has been commonly used in industrial production.Compared with the traditional manual detection,the machine vision system has higher detection accuracy and faster detection speed,which can greatly improve production efficiency.However,for some highly reflective surfaces in common machining processes,such as high-precision spindles and bearings,and electroplated surfaces,existing machine vision systems are unable to accurately identify surface defects due to excessive surface reflections that cause severe interference.The existing machine vision system for high-reflection surface defect detection has high design cost,complicated image preprocessing process,and it is difficult to extract reflective surface image features during image processing,and the calculation amount is large and the real-time performance is poor and limited.Therefore,manual detection is still being used for defect detection of highly reflective surfaces in some machining processes.Therefore,it is necessary to develop a new detection system with higher accuracy and more efficient detection of surface defects on highly reflective surfaces instead of manual detection to improve detection efficiency.In recent years,deep learning technology has developed rapidly,and convolutional neural networks have made computer vision research based on deep learning even further.Multi-layer convolutional neural networks excel in the field of image processing,which can effectively achieve image classification without complicated image preprocessing and manual image extraction.In this paper,a set of high-reflection surface defect detection system based on multi-layer convolutional neural network algorithm and ensemble learning algorithm is designed.The main work of this research includes:(1)A machine vision hardware system was designed and built according to the characteristics of the highly reflective surface to be tested.The reflected light model of the highly reflective surface and the possible optical noise are analyzed.According to the surface reflection irregular surface detected in this study,a machine vision device based on reflected light for image acquisition is designed.The parameters related to the camera and illumination are determined.Installation and commissioning enables efficient image acquisition of highly reflective surfaces.(2)Image acquisition is performed using a designed machine vision hardware system,and related image preprocessing algorithms are designed based on the acquired images.The data set is enhanced and the entire data set is labelled.(3)ConvPool-CNN-C model,ALL-CNN-C model,integrated learning model and SIFT feature point and perceptual hash algorithm are designed to classify images of highly reflective surface images to realize defect detection.For the three deep learning models,the structure,initialization method and optimization solution process are designed and trained using the training data set.The performance evaluation of the above method is carried out on the test set.The experiment shows that the accuracy of the surface defect detection of the high reflection surface is 97.76%,the recall rate is 98.8%,and the accuracy is 96.78%.The comprehensive evaluation index It is higher than the other three detection algorithms and has better comprehensive detection performance for highly reflective surface defects.(4)Characterize the highly reflective surface detected and explore the cause of the error.This paper theoretically analyzes the feasibility of migration learning methods for high-reflection surface defect detection applications and analyzes its application value based on several detection scenarios.