Abstract: Aiming at the problem that the traditional underwater object detector is greatly affected by the environment, a new lightweight network, called LUNet, is proposed to extract features, and a lightweight detector, called LUDet, is proposed by combining the two-stage detection algorithm. Firstly, in the first stage of the backbone network, efficient convolution pooling is used to obtain different feature expressions. Secondly, two-way dense connections are proposed on the basis of dense connection structure to improve the network representation ability. The network is composed of convolution pool layer and two dense connection structures. GhostModel is used to replace the 1×1 point convolution in the network. The classification experiments on CAFIR10 and CAFIR100 datasets show the effectiveness of the proposed backbone network. For the detection task, LUDet detects the target through feature maps obtained after channel attention and multi-stage fusion. The improved detection algorithm is validated using two underwater datasets. The mAP of the u nderwater biological dataset reaches 52.5%, and the mAP of the underwater garbage dataset reaches 58.7%.

Key words: underwater object detection, lightweight network, point convolution, two-way dense connection, channel attention