Detection And Classification Of Strip Surface Defects Based On Machine Vision

Due to many factors such as material specification,production equipment and rolling technology in cold rolling production,different types of defects are generated on the surface of the strip material,which affect the practicability and aesthetics.In this paper,the surface defect detection system is designed,the surface defect image library is established,the convolution neural network algorithm is improved,and the improved convolutional neural network is used to train and test the defect image library.First of all,for the problem of image acquisition and transmission of the surface defect detection device of the strip,two aspects of hardware and software are studied.In the hardware structure design,the light source,lighting mode and image sensor are discussed in detail.The FPGA processor and memory components such as SDRAM,ethernet DM9000 A and other peripherals are selected and the circuit connection problem is solved,and the Nios system is established on the FPGA.In the software design,the relevant register of CMOS sensor MT9M034 are configured in FPGA using the I2 C bus,and the MT9M034 acquisition program and DM9000 A IP core were prepared.The μC/OS-II real-time operating system and the Lwip protocol are introduced in the Nios to complete the system acquisition and transmission part.Secondly,according to the noise problem during the process of acquiring defect images,the noise types and common filtering methods are analyzed.A fast median filter based on FPGA is designed to implement fast median filtering of the defect images.Aiming at the lack of training samples in network training,the system’s image acquisition module was used to collect the surface defect images of the seven types of strips,and the images were manually extended to establish a surface defect image library of strips.In the end,the defect characteristics are extracted manually during the surface defect detection of the traditional strip,and the process is complicated.In view of this problem,a detection method based on convolution neural network was proposed,and the error rate of 6.86% was obtained by testing the defect image library on the strip surface.The convolution neural network is not high in the image library recognition rate because the feature extraction is not enough.Therefore,in this paper,Gabor convolution kernel is added to the network,and a deep convolutional neural network based on Gabor kernel is proposed.The low error rate of 0.28% was obtained by using the network to train and test on the surface defect image library.