Object Detection and Classification from Sonar Underwater Images using YOLO and DETR models

Underwater object detection using SONAR images is critical for marine conservation, environmental monitoring, waste management, and rescue operations. In this study, the authors apply various deep learning-based object detection models to the UATD SONAR dataset. The models include YOLOv5, YOLOv8, YOLOv10, YOLOv11, DETR and RT-DETR. All models were trained under consistent conditions to enable a fair comparison. Performance evaluation was conducted using mean Average Precision (mAP), precision, recall, and accuracy to assess the effectiveness and suitability of each model for underwater object detection tasks. The study found YOLOv8 to be the most effective deep-learning model for the defined task. To further enhance classification accuracy, a model ensembling strategy was implemented by combining the predictions of the best-performing models. This strategy achieved state-of-the-art results for the SONAR dataset.