YOLOv8-SC: an improved seafood target-detection model

The advancement of convolutional neural networks has significantly enhanced autonomous underwater fishing. This study introduces YOLOv8-SC, a lightweight model designed for detecting valuable seafood and sea cucumbers in aquatic robots with limited resources. YOLOv8-SC tackles the challenges of high parameter counts and memory usage in large architectures. It replaces the traditional C2f module with the Sequential Optimized Squeeze Excitation (SOSE) module to streamline the structure and reduce parameters. The model adopts a Bi-directional Feature Pyramid Network (BiFPN)-based architecture to improve accuracy without adding detection heads. The AGConv module and Compact Detect detector further reduce computational demands. Soft-NMS enhances detection in complex environments. Achieving a mean average precision (mAP@50) of 83.7% for seafood, YOLOv8-SC demonstrates strong accuracy with significantly lower resource requirements, making it well-suited for underwater detection tasks.