Design Of Industrial Image Classification And Detection System Based On Deep Learning

With the continuous improvement of industrial production efficiency,the testing standards for industrial products are becoming increasingly stringent.Traditional image processing methods are usually based on pre-designed feature extractors,which optimize the detection results by constantly adjusting parameters.The detection effect of industrial images with complex texture is poor.In contrast,deep learning can better represent the essential information of the data set by training a large number of samples,and is highly demanded to classification and detection of surface defects in industrial practice.It provides an innovative solution to the problems of high false detection rate and low accuracy of the current traditional algorithms.In this paper,an industrial image classification and detection algorithm based on GAN and Faster R-CNN and an industrial image classification and detection algorithm based on YOLOv2 are proposed.The research mainly includes the following aspects:(1)The problem of industrial image classification and detection is studied,and an industrial image classification and detection system based on deep learning is designed.The system mainly includes four subsystem: image acquisition,motor drive and control,data processing and human-computer interaction,and robot coordinated control.After taking fabric images with cameras,it judges whether the fabric contains defects,and then records the location information and types of defects.The proposed system realizes real-time classification and detection of defects for industrial images.In the output control part,the robot technology is introduced to sort the defective industrial products through the robot's mechanical arm,so as to improve the quality of industrial products.(2)Industrial image classification and detection algorithm based on GAN and Faster R-CNN is proposed.Aiming at the problems of uneven distribution of defect samples,insufficient number and type of defect samples in the process of industrial image acquisition,this paper firstly extracts the potential features and probability distribution of defect by GAN,then extracts defect contour based on candidate box generation algorithm,and learns its regional and edge features by deep convolution neural network.Finally,the non-maximum suppression algorithm is used to adjust the optimized defect position,and the Softmax classifier classifies the defect categories.The average accuracy can reach about 94.93% when validated on the industrial image data set.(3)To address the problem of real-time classification and detection of industrial images,fabric defect detection based on YOLOv2 deep learning model is proposed.In the classification and detection of defects for industrial images,YOLOv2 algorithm is used to predict the location and types of defects directly.YOLOv2 algorithm introduces local weight sharing mechanism to enhance the positioning performance of the network.It also refers to advanced algorithms such as full convolution neural network and knowledge extraction,and adopts integrated detection scheme are also introduced to extract the edge features of defects directly from the original image,which greatly reduce the computational complexity of industrial image processing and successfully improve the industrial image classification based on deep convolution neural network.And the real-time performance of the detection algorithm.Experiments prove that YOLOv2 object detection network can achieve satisfying detection results on industrial image data sets.The average accuracy of classification and detection is about 98.18%.At the same time,it can meet the real-time requirements of the industrial site.There are 43 figures,12 tables and 62 references in this paper.