MCMS-Net: an underwater target detection network based on Multi-Conv Multi-Scale feature fusion

Influenced by the complex characteristics of the marine environment, a Multi-Conv Multi-Scale Network (MCMS-Net), a multi-scale feature fusion network for underwater target detection, is proposed to address the problem of poor underwater target detection accuracy due to the difficulty of capturing target features of marine organisms. Firstly, a backbone network MCM Backbone stacked by Multi-Conv Module1 (MCM1) and Multi-Conv Module2 (MCM2) is designed to improve the amount of feature residual information of the target to be detected in the network, which is conducive to the subsequent feature transfer. Secondly, the SE attention mechanism is introduced to improve the network’s ability to capture shallow and high-level feature information through the MCM2-Path Aggregation Network (MCM2-PANet) with multi-scale feature fusion. Finally, an anchor frame-based MCM Detection Head is designed to improve the detection accuracy of different targets by combining multiple sets of convolutional module operations with different sized feature layers. Experiments show that the algorithm is able to reach 82.8% detection accuracy on underwater target URPC dataset, which is informative for target detection applications of underwater robots in real engineering.