An Improved YOLO11n-Based Method for Sea Cucumber Detection in Complex Underwater Scenes

Abstract

Sea cucumber detection in underwater scenes is challenging owing to low contrast, blurred boundaries, and interference from visually similar organisms. This paper proposes an improved YOLO11n detector that incorporates RFCAConv for position-sensitive feature extraction, an Edge Information Enhancement Module (EIEM) for boundary-aware fusion, and Adaptive Threshold Focal Loss (ATFL) for hard-sample emphasis. A single-class sea cucumber dataset is constructed from the DUO benchmark. On the validation set, the proposed model improves mAP@0.5:0.95 from 66.3% to 69.4% over YOLO11n. Three independent test-set runs with random seeds 42, 3407, and 2025 yield AP@0.5 of $76.2 \pm 0.3 \%$ and mAP@0.5:0.95 of $54.3 \pm 0.3 \%$, with all three seeds outperforming the baseline, representing modest but repeatable gains. On a hard subset containing confusing sea urchins, scallops, and starfish, confusing-object false positives are reduced from 37 to 24, and on a hard-negative subset, false positives decrease from 32 to 21.