Real-time detection of small underwater organisms with a novel lightweight SFESI-YOLOv8n model

To address the challenges of detecting small targets in complex underwater environments, an efficient and lightweight model, SFESI-YOLOv8n, is proposed. The model improves small target recognition by incorporating a dedicated detection layer and reduces parameter count by removing the large target detection layer. Furthermore, the introduction of the C2f-F module eliminates redundant information from consecutive convolution operations in the bottleneck, further simplifying the model. The integration of a lightweight mixed local context attention (MLCA) mechanism within the small target fusion layer increases sensitivity to small targets. The dynamic upsampler (DySample) employs point sampling to preserve enhanced edge and detail information in feature maps, resulting in clearer feature representations. In addition, the novel In-NWD loss function, utilizing Wasserstein distance and auxiliary bounding boxes, improves small target detection performance. On the UPRC2020 public dataset, SFESI-YOLOv8n achieved an mAP@0.5 of 83.7%, which is a 1.1% improvement over the baseline model. The parameter count and size of the model were reduced by 49.2% and 45.7%, respectively. The frame rate reached 227 FPS, indicating a 9 FPS increase compared to the baseline. On the NVIDIA Jetson TX2 edge device, inference latency decreased from 65 ms to 24 ms with TensorRT acceleration, thereby meeting real-time detection requirements. The SFESI-YOLOv8n model provides a viable and efficient solution for the autonomous detection of small underwater targets, demonstrating significant practical value.