The steel industry is a pillar industry of the national economy, accounting for about 5% of China's GDP. As one of the important raw materials in the iron and steel industry, the surface quality of strip steel directly determines the quality of the final product. However, in the actual production process, due to the limitations of the process flow and production environment, the surface of the strip steel often appears scratches, punching, silk spots, inclusions and other defects. These defects not only affect the appearance quality of the product, but also may cause the steel to be more prone to corrosion and rupture, thereby affecting the safety and reliability of the product. The traditional strip defect image detection system is difficult to meet the needs of millimeter real-time processing because of its limited processing speed. In contrast, FPgas have powerful parallel processing capabilities and low latency characteristics, which are particularly suitable for applications such as strip surface defect detection requiring real-time feedback and high precision. Therefore, whether in order to ensure the stable supply of the domestic market, or in response to the national "Belt and Road" initiative to promote foreign exports, the FPGA-based strip surface defect image detection system has important theoretical significance and broad practical application prospects. According to the requirements of strip surface defect detection, Intel Cyclone IV EP4CE10F17C8 is selected as the FPGA chip, OV5640 CMOS camera is used for image acquisition, and W9825G6KH-6 SDRAM is used for image data storage. PCF8591T is used for AD/DA conversion of video data and successfully builds a complete hardware platform. The platform realizes the whole process from the acquisition, processing, storage, transmission to the final display of the strip defect image. Implement the FPGA top-down design process, first of all, design the IIC communication protocol interface, which is used to configure the camera and realize its acquisition function. According to the SDRAM manual, the initialization, automatic refresh, read and write operations and arbitration commands of SDRAM are completed in order to efficiently store and manage image data. In order to coordinate the reading and writing of image data better, an asynchronous FIFO control module is designed to ensure the stability and high efficiency of data transmission. Finally, based on VGA sequence diagram, VGA driver circuit is designed to realize the real-time display of image data on the display screen. In view of the high sensitivity of traditional Canny algorithm in the face of salt-and-pepper noise, an adaptive median filter is designed to replace the Gaussian filter. By dynamically adjusting the size and form of the convolutional template, the method can adaptively select the most suitable filtering mode according to different noise levels, effectively suppress the interference of salt and pepper noise, and improve the image quality. In order to solve the problem of false detection that may occur in the recognition of edge points by the traditional Canny algorithm, this paper uses a four-direction 3×3 Sobel operator to replace the original two-direction 2×2 convolution template. This improvement enables edge detection in four directions: horizontal, vertical, top left to bottom right, top right to bottom left, making edge detection more comprehensive and clear. In addition, to solve the problem that the uncertainty of image display may be caused by manually setting the threshold in the traditional Canny operator, an adaptive threshold method is proposed, and the relationship between high and low thresholds is set to double. Finally, the system test and analysis are carried out. Through the comprehensive comparison of each filter algorithm template, the results show that the adaptive median filter algorithm has the best performance in both subjective visual effect and objective indicators PSNR and SSIM, which is suitable for this system. Combined with the joint simulation of MATLAB and ModelSim, the improved algorithm is superior to the traditional method in terms of subjective visual effect, and can capture image edges more accurately. Moreover, the real-time transmission effect in the on-board experiment is good, and it can support real-time image processing and defect detection tasks. System resource consumption is low, the total time of FPGA processing a 640×480 resolution image is 13.03 milliseconds, 10 times faster than the software algorithm