A Method for Extracting Planar Image Features based on Convolution Neural Network

This paper is supported by the 2022YFG0070 science and technology plan of Sichuan Province. Image feature extraction technology plays a very important role in the measurement and control process of industrial product processing, but for planar micro defects in the background, the current general single shot multibox detector (SSD) has some disadvantages, such as easy loss of feature information, low detection accuracy, and insufficient number of detection feature maps. In view of the above problems, combined with the characteristics and requirements of image feature extraction in the process of measurement and control processing, this paper proposes and designs BSSD algorithm. The algorithm uses ResNet34 to extract more micro defect information to solve the problem of feature extraction; Seven multi-scale feature maps were selected to increase the number of feature maps for detecting micro defects; A backtracking layer is set up to fuse the abstract information of the high-level network into the shallow network before the multi-scale feature map is input into the classification network to enhance the expression ability of the abstract features. The experimental data show that the accuracy is comparable to DSSD, and the speed is similar to FSSD, which shows a significant advantage in the accuracy of small target detection.