Marine Fish Object Detection Based on YOLOv5 and Attention Mechanism

Abstract

Automatic fish object detection methods based on computer vision have been widely used in recent years. In this paper, a YOLOv5s marine fish detection method that integrates attention mechanisms is proposed, which has alleviated the issue of misdetections and omissions caused by the shape, color, and size of the marine fish object. First, the multi-scale clustering method is used to recalculate the anchor box for matching the dataset. Then, the channel attention module is added to the feature extraction network of YOLOv5s by using the lightweight attention ECA-Net to obtain important feature information. Meanwhile, the spatial attention module SSA is used to enhance the edge feature information and improve detection accuracy. To further increase the generalization of the detection model, the DropBlock regularization method is added to sparse the features extracted from the backbone network to lessen the overfitting problem. The network is trained and tested using a combination of two publicly available datasets and a self-built dataset with rich samples covering complex living environments of marine fishes. Experiments demonstrate that our method can effectively increase the accuracy of fish object detection while also maintaining the speed of object detection.