<![CDATA[Underwater object detection is an important application of computer vision for automated monitoring systems. However, detection performance in underwater environments is hindered by unstable illumination, image blur, and optical distortion from refraction, which degrade visual quality and feature representation. Existing attention-based approaches can enhance feature extraction and detection performance, but their additional computational complexity may limit inference efficiency and deployment on resource-constrained devices. To address these limitations, this study proposes a lightweight modification of YOLO12-nano that combines channel tuning with a newly designed Enhancement Channel Recalibration (ECR) module. Unlike previous attention-based approaches, ECR is specifically designed to enhance the extraction and recalibration of channel-wise information, enabling more effective feature refinement without introducing substantial computational overhead. In addition, channel dimensions within the YOLO12-nano architecture are reduced to obtain a more lightweight detector while preserving its detection capability. Experimental results on the Real-World Underwater Object Detection dataset demonstrate that the proposed detector achieves an mAP50-95 of 61.8%, representing a 1.31% improvement over the original YOLO12-nano, while reducing the number of parameters by 22.12% to 2.0 million and computational cost by 3% to 6.3 GFLOPs. The proposed model also achieves an inference speed of 23.08 FPS on a CPU, compared with 21.3 FPS for YOLO12-nano. These results demonstrate that the proposed combination of channel tuning and ECR provides a favorable balance between detection performance and computational efficiency, making the detector more suitable for real-time underwater monitoring and deployment on resource-constrained devices.]]>
Copyrights © 2026
