A YOLOv5s model improvement method

The current positioning detection models have problems such as large size and low detection efficiency when applied to low computing edge devices, which limit the deployment and performance of subsequent control algorithms. In this paper, based on three typical deep learning models: the Faster R-CNN, SSD, and YOLOv5s, the YOLOv5s object detection model with stronger detection performance is identified through a migration learning approach using the COCO2017 public dataset for training and performance evaluation. To solve the problems of low detection efficiency and real-time response rate when deep learning algorithms are applied on embedded devices, this paper refers to the MobileNet network to improve the YOLOv5s backbone network, and enhances the YOLOv5s necking network by utilizing the C3Ghost module composed of Ghost Bottleneck. After model training and performance evaluation, the results show that the constructed MG-YOLOv5s lightweight model has better performance, which is important for the deployment of deep learning algorithms on embedded devices with low arithmetic power.