Research on object recognition algorithms for underwater visual degraded images

Addressing the degradation in performance of conventional detection algorithms caused by low image contrast, blurred features, and complex backgrounds in underwater turbid environments, this paper proposes the ERCS-YOLO model. Built upon the YOLOv11 architecture, it represents an efficient object detection method tailored for turbid underwater settings. This model incorporates advanced modules including EIEStem, RGCSPELAN, CGRFPN, and SEAMHead. By enhancing multi-scale feature fusion capabilities through an improved feature extraction structure, and integrating context-aware and spatial attention mechanisms, it significantly boosts the model's feature representation capacity under degraded conditions. Concurrently, optimised pre-processing and post-processing workflows achieve synergistic improvements in both accuracy and efficiency. Experimental validation on a self-constructed underwater turbidity target dataset demonstrates the proposed model achieves an accuracy of 0.872, surpassing the baseline YOLOv11n by 0.018. The mAP50 reaches 0.8987, representing a 2.75% improvement, while mAP50-95 attains 0.6149, showing a 1.9% enhancement. Experimental results demonstrate that this method effectively enhances detection accuracy and robustness for targets in turbid underwater environments.