Efficient detection method of deep-sea netting breakage based on attention and focusing on receptive-field spatial feature

Fish escapes due to breaches in deep-sea netting can affect local ecosystems. To accurately and quickly detect broken netting, we propose YOLOv7-net, an efficient deep-sea netting breakage detection method based on attention and focusing on the receptive-field spatial feature. First, Bi-level Routing Attention (BRA) is introduced to enhance the acquisition of feature information at different scales. Second, a new coordinated attention module (RFCAConv) that focuses on the spatial features of the receptive field is used to capture more detailed feature information. Finally, a new network module called CFE that integrates efficient channel attention (ECA) and FasterNet during cross-stage connections is designed, enhancing the ability of the network to express features while reducing the number of required parameters and computational complexity. The results obtained on a self-constructed broken netting dataset show that the precision, recall, AP, F1 score and detection speed of YOLOv7-net are 2.8%, 1.8%, 2.4%, 2%, and 8.92 fps higher than those of YOLOv7, respectively, and the proposed approach can be specifically used to identify deep-sea netting damage. Our method improves the efficiency of broken netting detection in complex marine environments, providing new insights into the development of mariculture equipment and the protection of ecosystems.