LMOD-YOLO: A lightweight YOLO model for marine organism visual detection

Abstract

Abstract Real-time detection of marine organisms plays a critical role in underwater ecological monitoring, endangered species protection, and autonomous underwater vehicle (AUV) operations. However, the degraded underwater images with low contrast and detail blur and limited embedded computing resources make it challenging for existing methods to balance accuracy and real-time performance. To address these problems, this paper proposes LMOD-YOLO, a lightweight real-time marine organism detection model. The whole model integrates three key modules to improve computational efficiency while maintaining high detection accuracy. Cross Stage Local Partial Convolution (CSLPC) module replaces the original C2f module to reduce redundant computation, improving spatial feature extraction efficiency. An Adaptive Spatial-Channel Dual Convolution enhanced Cross-scale Feature Fusion module, termed ASCDC-CCFM, is designed in the neck to achieve efficient and lightweight fusion of multi-scale features from the backbone network. Furthermore, the Reparameterized Detection Head (RepHead) decouples classification and regression branches during training and reparameterizes them into a single convolution layer during inference, improving detection performance on degraded images, cluttered backgrounds, and dense small targets. Experiments conducted on the CUDD dataset show that LMOD-YOLO reduces computational cost and parameter count by 51.1% and 59.7%, respectively, and improves inference speed by 36.9% compared with YOLOv8m, with only slight decreases of 0.006 and 0.012 in mAP@0.5 and mAP@0.5:0.95.Deployment on an NVIDIA Jetson AGX Orin achieves a mean inference latency of 2.90 ms and an energy efficiency of 11.76 FPS/W. These results demonstrate a favorable balance between detection accuracy and deployment efficiency, enabling timely responses in dynamic underwater monitoring while potentially reducing the energy burden of battery-powered AUVs. The source code is available at https://github.com/yangxyyy/LMOD-YOLO .