Torpedo detection algorithm based on improved YOLOv5s

DOI:	10.11809/bqzbgcxb2024.01.005
Keywords:	torpedo detection; YOLOv5s; deep learning; separable vision transformer; attention mechanism; bidirectional feature pyramid network
Abstract:	Aiming at the problems of low detection accuracy and slow detection speed in deep sea torpedo detection, a torpedo detection algorithm based on improved YOLOv5s is proposed.The separable Vision converter (SepViT) block is used to replace the C3 modules in the last layer of the backbone network to enhance the connection between the backbone network and the global information, to extract the torpedo features, and to reduce the omission rate and error detection rate.ECA attention mechanism is introduced into the backbone layer network of YOLOv5s network model to improve the extraction ability of the detection model for torpedo deep level key features in the complex deep sea environment, avoid dimension reduction, and capture cross channel interactive information in an effective way to improve the detection accuracy of the torpedo detection model. The Path Aggregation Network (PANet) in the neck layer of the network model is replaced by the Bidirectional Feature Pyramid Network (BiFPN), and cross scale connection is used to remove the nodes in the Path Aggregation Network (PANet) that have less contribution to feature fusion to achieve rapid fusion of multi scale features and to improve the detection efficiency of the torpedo detection model. The experimental results show that the mean average accuracy (mAP) of the improved YOLOv5s torpedo detection algorithm reaches 97.0%, which is 3.7 percentage points higher than the original YOLOv5s algorithm, and the detection speed reaches 83 FPS, which effectively improves the accuracy and speed of deep sea torpedo detection.
Published:	2024-01-28
PDF