LMC-YOLO: a lightweight underwater benthic organism detection network with multi-scale feature extraction
摘要
With the advancement of science and technology, underwater robots have become the standard means for harvesting benthic marine organisms. Accurate identification of these organisms is a critical prerequisite for the proper functioning of such robots. However, most existing object detection models are too computationally intensive to be fully deployed on resource-constrained underwater platforms. To address this issue, we propose a lightweight object detection network tailored for benthic marine organisms, named LMC-YOLO, which incorporates multi-scale feature extraction. To begin with, we introduce a lightweight and efficient feature extraction module designed to enhance the model’s ability to capture salient features while reducing its overall complexity. Additionally, we propose a multi-branch semantic enhancement module that leverages branches with diverse receptive fields to capture the multi-scale characteristics of benthic marine organisms. Furthermore, we incorporate a channel-spatial attention mechanism to improve the model’s capacity to distinguish semantic information and increase sensitivity to critical features. Finally, we apply pruning techniques to reduce both the computational load and the number of parameters in the model, enabling more efficient deployment on embedded systems. Experimental results on the URPC data set demonstrate that, compared to the baseline model YOLOv7s, the proposed LMC-YOLO achieves a 3.2