Skip to content
Open access

An Adaptive YOLO Framework for Coral Reef Object Detection With DUAA, MLCA, and CAL

2026 · IEEE Access · Vol 14, pp. 156571-156591 · 0 citations · 45 references

TL;DR

YOLOv12n is enhanced with Dynamic Underwater Adaptive Attention, Multi-Level Channel Attention (MLCA), and a Coral-Adaptive Loss (CAL) to enhance feature refinement and multi-scale representation in complex underwater environments.

Abstract

Underwater object detection is intrinsically challenging owing to scale variations, visibility degradation, and variable illumination that challenge conventional deep learning models. We enhance YOLOv12n with Dynamic Underwater Adaptive Attention (DUAA), Multi-Level Channel Attention (MLCA), and a Coral-Adaptive Loss (CAL) to enhance feature refinement and multi-scale representation in complex underwater environments. DUAA dynamically adjusts spatial and channel-wise attention using image-quality-aware modulation, while MLCA performs lightweight channel recalibration on hierarchically fused multi-scale neck features using global average pooling and a two-layer MLP. CAL integrates GIoU localization, adaptive focal classification, IoU-aware confidence, and bounding-box-guided attention supervision into a task-specific objective, with its weights selected through Bayesian optimization. Experiments on the Coral Reef Bleaching v16 Balanced dataset show that the full DUAA–MLCA–CAL model achieves a test-set mAP@50 of 0.9191, compared with 0.9074 for YOLOv12n (1.29% relative improvement), with precision of 0.8896, recall of 0.9030, mAP@50–95 of 0.7729, and F1-score of 0.8962. The model contains 2.5385 M parameters and 3.0045 GFLOPs, achieving 41.24 FPS on a Tesla T4 and 56.07 FPS on an NVIDIA RTX 3070 under batch-one FP32 inference. Zero-shot inference on URPC2019, Aquarium, and Brackish characterizes out-of-domain prediction behavior, while cross-version evaluation on Coral Reef Bleaching v13 evaluates robustness to dataset-version shift rather than independent quantitative generalization.

Read PDF

Similar papers

Open access Sep 2026

Underwater tiny object detection network based on multi-scale attention and adaptive feature fusion

Experimental results demonstrate that the proposed Bidirectional weighted Concat with Efficient multi-scale Attention You Only Look Once (BCEA-YOLO) method outperforms competing methods for underwater object detection, and edge deployment tests on the NVIDIA Jetson platform validate its real-time inference efficiency.

Xing-Yu Wang, Yu-Han Lin, Quan J. Wang et al. · 0 citations
Open access Sep 2026

YOLO-CBNet: A robust attention-enhanced detection framework for underwater fish recognition in aquaculture environments

The accurate identification of fish species remains a pivotal challenge in aquaculture, as it directly influences population management, health monitoring, and overall production efficiency. Yet, underwater environments often present difficult conditions such as reduced visibility, suspended particles, and uneven light...

Mahdi Hamzaoui, Leila Bousbia, Mohamed Ould-Elhassen Aoueileyine et al. · 1 citation
Open access Sep 2026

FH-YOLO: a cross-stage stabilized framework for underwater benthic object detection

Both quantitative and qualitative results demonstrate that the proposed cross-stage coordination significantly improves detection accuracy and robustness in complex underwater environments.

Meng-Ya Ma, Zhang-Wei Yu, Chan-Gan Ren · 0 citations
Conference Sep 2026

Improved YOLO11s-Based Net-Hole Detection for Deep-Sea Aquaculture Platforms

Net-hole detection is critical for the structural safety of deep-sea aquaculture platforms, yet automated inspection faces challenges from extremely small damage regions, complex underwater backgrounds, and variable illumination. This paper proposes three improvements to YOLO11s: embedding a Convolutional Block Attenti...

Wen-Qiang Liu, Cheng-Qiang Yu, Wei He et al. · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.