LT-YOLO: A Lightweight and Real-Time Underwater Target Recognition Method Based on Staged Convolution and Channel Attention

The main innovative achievements of this paper are as follows:

(2) Addressing the issue of insufficient real-time performance caused by target recognition networks relying excessively on network depth or complex architectures, a lightweight real-time underwater target recognition method called LT-YOLO is proposed by analyzing the role of each module in YOLOv5s and its impact on network complexity. Firstly, YOLOv5s is reconstructed by leveraging the stage-wise feature extraction capability and low-rank approximation characteristics of the Ghost module, as well as the strong feature representation ability of depth-wise separable convolution. This reconstruction reduces the model parameter count while enhancing the network's perception of fine-grained information. Then, a channel self-attention module is introduced to enhance the interaction between different feature channels and improve the network's expression of effective features. Experimental results on popular databases like RUOD, self-built database UCOD, and pool experiments demonstrate that LT-YOLO achieves consistent recognition accuracy and robustness compared to YOLOv5s, with a speed improvement of approximately 3.4 times on embedded processors.