LFN-YOLO: Precision underwater small object detection via a lightweight reparameterized approach

Mingxin LIU ^1,2 Yujie Wu ³ Li RuiXin ⁴ Cong Lin ^1,2*

• ¹ School of Electronic and Information Engineering, Guangdong Ocean University, Zhanjiang, Guangdong Province, China
• ² Guangdong Provincial Key Laboratory of Intelligent Equipment for South China Sea Marine Ranching, ,Zhanjiang, China
• ³ Guangdong Ocean University, College of Naval Architecture and Shipping, Zhanjiang, China
• ⁴ College of Naval Architecture and Shipping, Guangdong Ocean University, Zhanjiang, China

Underwater object detection plays a significant role in fisheries resource assessment and ecological environment protection. However, traditional underwater object detection methods struggle to achieve accurate detection in complex underwater environments with limited computational resources. This paper proposes a lightweight underwater object detection network called LightFusionNet-YOLO (LFN-YOLO). First, we introduce the reparameterization technique RepGhost to reduce the number of parameters while enhancing training and inference efficiency. This approach effectively minimizes precision loss even with a lightweight backbone network.Then, we replaced the standard depthwise convolution in the feature extraction network with SPD-Conv, which includes an additional pooling layer to mitigate detail loss. This modification effectively enhances the detection performance for small objects. Furthermore, We employed the Generalized Feature Pyramid Network (GFPN) for feature fusion in the network's neck, enhancing the network's adaptability to features of varying scales. Finally, we design a new detection head, CLLAHead, which reduces computational costs and strengthens the robustness of the model through cross-layer local attention. At the same time, the DFL loss function is introduced to reduce regression and classification errors. Experiments conducted on public datasets, including URPC, Brackish, and TrashCan, showed that the mAP@0.5 reached 74.1%, 97.5%, and 66.2%, respectively, with parameter sizes and computational complexities of 2.7M and 7.2 GFLOPs, and the model size is only 5.9 Mb. Compared to mainstream vision models, our model demonstrates superior performance. Additionally, we conducted simulation experiments by deploying LFN-YOLO on the NVIDIA Jetson AGX Orin edge computing device. The results demonstrate that LFN-YOLO achieves higher real-time performance and superior suitability for underwater applications, further indicating the exceptional capabilities of our model.