关键词: 目标检测, ＹＯＬＯｖ４, ＴｅｎｓｏｒＲＴ, 嵌入式平台

Abstract:  This paper proposes a lightweight broad bean sprout detection method based on the improved YOLOv4 network.The original backbone network CSPDarknet53 of YOLOv4 is replaced by the MobileNet network, and, after it is replaced by depthwise separable convolution, ordinary convolution of the feature extraction network and the prediction layer is strengthened. After being improved by network training, the network structure is reconstructed and optimized by using NVIDIA acceleration engine TensorRT, the GPU operation efficiency improves, and inference and acceleration of the model on the embedded platform are realized. The experimental results show that the improved network volume reduces to approximately 20% of the original volume, with a reduction of AP by only 3.14%, but the detection speed is 4 times that of the original network. On the Jetson Nano embedded platform, the inference speed of the improved network model reaches 20.3 FPS, which shows the proposed network model can contribute to the real-time application of the deep learning model on the embedded platform.