부산대학교 도서관

In this paper, we design a circuit components testing system for physics teaching based on embedded devices. In order to solve the problem that the computing performance of embedded devices based on NPU processor is weak and the same class samples are not detected due to high overlap, ReD_SSD Mobilenet algorithm based on SSD algorithm is proposed. Depth separable convolution is used to replace the traditional convolution of SSD algorithm to reduce the computational complexity of the model; ReDistance-IOU algorithm is used to complete the screening operation of the model pre-detection frame to reduce the system undetected rate; At the same time, the off-line 8 bit mixed quantization is adopted for the trained network model to further reduce the computational complexity of the model within the acceptable range of accuracy loss. The experimental results show that, compared with the SSD algorithm model, the ReD_SSD Mobilenet algorithm model reduces the average missing rate by 37. 4% on the physical circuit component data set. After adding the depth-separable convolution and off-line 8-bit mixed quantization, the compression rate of the model reaches 95%. The results show that the detection accuracy of ReD_SSD Mobilenet algorithm reaches 76.2%, and the detection frame rate on an embedded device is as high as 33.4 frames/s.