Real-time detection of underwater objects in Side Scan Sonar Data using Deep Learning Models

Accurate identification, classification and localisation of underwater objects in side-scan sonar images is one of the active areas of research. In order to deploy deep learning-based object detection models in resource-constrained embedded systems onboard autonomous underwater vehicles, the models must be optimised to reduce the model size and the inference time to meet real-time requirements. This paper details the works carried out in dataset generation, model training, testing and evaluation on side-scan data for real-time inference. The side-scan sonar data is generated using a remotely operated vehicle fitted with side-scan sonar. The objects of interest, such as cylindrical and truncated conical-shaped objects are deployed in the tank and harbor for scanning with imaging sonar. The sonar images are pre-processed, augmented and annotated for training deep learning-based object models. Popular generic object detection models YOLOv5 and YOLOv8 have been finetuned and trained on a custom side scan sonar dataset. The trained model is quantized and ported onto the custom-embedded hardware for inference in real-time. Our experiments show that both YOLOv5 and YOLOv8 are performing well on custom-side scan sonar dataset. However, the YOLOv5 large model showed better detection accuracy and inference time performance. The training mAP of $66.7 \%$ for YOLOv5L and $66.5 \%$ for YOLOv8L is achieved. A frame rate of 24FPS for YOLOv5L and 16FPS for YOLOv8L is achieved after 16-bit quantization. The models have been evaluated on NVIDIA Xavier NX SOM-based custom-embedded hardware board. As part of this study, it is established that both YOLOv5L and YOLOv8L models are suitable for real-time implementation and deployment on AUVs.