Underwater Cultured Sea Cucumber Recognition Method Based on Multi-Strategy Enhancement

Highlights In this study, a multi-strategy enhanced recognition method for underwater sea cucumber based on YOLOv8n, DHGY, was designed. The automatic color level algorithm is used for image enhancement, the dynamic snake convolution and multi-scale feature fusion structure are combined to improve the boundary sensitivity and cross-scale feature extraction ability. The lightweight design is introduced to reduce the computational complexity. Abstract. In the process of sea cucumber identification, the low visibility of the underwater environment and complex background interference lead to significant challenges in the precision detection and positioning of sea cucumber targets. In addition, the general object detection model is difficult to meet the needs of underwater robots for real-time operation due to its high computational complexity. To address the aforementioned issues, this study proposes an identification method for underwater-cultured sea cucumbers utilizing a multi-strategy enhancement approach. Firstly, to address the issue of low underwater visibility, the automatic color level algorithm is employed for color correction, and the optical attenuation of the water body is compensated by dynamically adjusting the pixel distribution, thereby enhancing the data quality of the detection model. Secondly, the dynamic snake convolution and HS-FPN multi-scale feature fusion structure are introduced into the detection network to enhance the detection precision and adaptability of the model to underwater targets at different distances, thereby improving boundary sensitivity and cross-scale feature extraction ability. Finally, to address the high computational complexity of the general model, this article adopts the GhostNet lightweight network structure and utilizes the feature mAP@50 redundancy elimination mechanism to significantly reduce computational cost while ensuring detection precision, thereby effectively meeting the application requirements of underwater real-time detection. The experimental results show that the optimized model, while maintaining lightweight characteristics, improves detection precision by 0.7%, mAP@50 by 4.2%, reduces the model size and parameter count by 30%, and increases inference speed by 37.8 fps, demonstrating superior detection performance and robustness compared to the baseline model.