Enhanced Faster R-CNN for Real-Time Underwater Object Detection

Detecting objects in underwater environments remains a critical challenge due to factors such as low visibility, noise, and severe image distortion. These challenges hinder the efficient monitoring of marine biodiversity and sustainable resource management, which are essential for addressing global ecological and economic concerns. This research proposes a novel approach to underwater object detection by enhancing the Faster R-CNN framework with several state-of-the-art techniques. Key innovations include the integration of Res2Net101 for multi-scale feature extraction, Generalized Intersection over Union (GIoU) for bounding box regression, and Soft Non- Maximum Suppression (Soft-NMS) to refine detection results. Additionally, Online Hard Example Mining (OHEM) addresses class imbalance during training, while mosaic data augmentation and multi-scale training improve model robustness and accuracy. The proposed model improves performance in challenging underwater environments, offering precise real-time object detection capabilities. This work has significant implications for marine ecological research, environmental monitoring, and the sustainable management of aquatic resources. By advancing automated detection systems, the study supports efforts to mitigate human impact on marine ecosystems and contributes to conservation initiatives.