Research On Product Surface Quality Detection Based On Deep Learning

At present,the artificial detection method is still widely used in most industrial production,but with the development of society,the industrial transformation is "high-end",the aging problem of population is increasingly serious,the human cost is rising,and the human eye detection method is increasingly unable to meet the requirements of product quality and production efficiency.Human eye detection is influenced by many objective and subjective factors,such as the level of knowledge and skills,working environment and fatigue degree,which can not guarantee the standardization and stability of detection results.In the visual detection based on the object detection technology,the surface image of the product is collected by the optical imaging method and detected by the object detection technology.Compared with the artificial detection method,the object detection technology is not limited by subjective and objective conditions,and can achieve non-contact fast online detection.In product surface detection,object detection is the most important research direction.Object detection refers to pointing out the possible location of the object such as defect in the collected product surface image,and providing more complete category location information.Because it is time-consuming and labor-intensive to perform detection completely by humans,the visual detection method based on object detection has high speed and good stability,and has been paid special attention by the industrial system development in recent years and has been widely used.There are two main research methods of product surface object detection: the method of object detection based on artificial design features combined with machine learning classifier and the method of object detection based on self extracting features of deep learning.Because the artificial design feature requires the designer to have rich prior knowledge,and the algorithm based on the artificial design feature is difficult to be extended to other related applications or repeated use,so aiming at the detection efficiency and detection accuracy in the complex industrial environment of the product surface image,a object detection method based on the deep learning self extracting feature is proposed.The main research contents are as follows:(1)For the object detection method of self-extracting features of deep learning,based on the improved SSD regression object detection algorithm,a lightweight object detection network IV-Mobile Net-SSD detection algorithm is proposed.IV-Mobile Net-SSD uses Mobile Net’s main structure depth separable convolution to redesign the backbone feature extraction basic network and replace the feature layer in the basic network with an Inception structure,which greatly reduces the model’s computational consumption and balances the contradiction between detection accuracy and detection speed.(2)Further considering the loss of small object detection by IV-Mob ile Net-SSD,the global context information feature map generated by cascading LSTM-Attention is used for detection,and an IVC-Mobile Net-SSD detection algorithm for multi-type feature map cascading is proposed.In addition,the K-means mean clustering algorithm is used to mine the aspect ratio parameters of the template box of the data set to improve the detection accuracy.The loss function is improved to accelerate the model training convergence.Experiments show that the network convergence speed is accelerated,and the detection performance of small objects has been improved.(3)Based on the machine vision method,this paper establishes a real-time detection system for mobile phone polarizer based on Windows platform.The system includes four modules: image acquisition,image preprocessing,polarizer surface quality detection and software interface design,which basically meets the design function requirements.The test results show that the single image detection process takes 200 ～ 700 ms and the detection accuracy rate is as high as 85.1%.