A high-precision and lightweight underwater fish detection and recognition approach based on the improved YOLOX-nano algorithm

In the past few years, the aquaculture industry has experienced a growing demand for in-situ fish detection using live videos from underwater observatory platforms to autonomously monitor cultivated fish and assess their health and growth conditions. To achieve the ideal balance between model size and detection precision, this paper offers a lightweight YOLOX-based fish detection and recognition approach termed Foc_YOLOXn_ASFF. In this study, we introduce Focal loss as a substitute for the binary cross-entropy (BCE) loss utilized by the original YOLOX to train the objectiveness branch, aiming to alleviate the class imbalance between the background and foreground classes. Meanwhile, we focus on moderating the inconsistency across various feature scales by combining YOLOX-nano with the adaptively spatial feature fusion (ASFF) structure to better fuse multi-scale features. Foc_YOLOXn_ASFF remarkably boosts the AP by over 3.6 % from 0.846 to 0.877 compared with the original YOLOX-nano through the implementation of Focal loss and ASFF structure. Foc_YOLOXn_ASFF achieves superior performance with 0.877 AP but only 2.27 M parameters, outperforming the counterparts YOLOv5n and SE_YOLOv5s_DGhost by 1.2 % AP and 7.9 % AP respectively. Extensive experimental findings indicate that Foc_YOLOXn_ASFF for both model size and detection precision is capable of meeting the requirements of practical applications on embedded devices with limited computational resources.