Comparison of Deep Learning Strategies for the Identification of Artificial Objects in Underwater Environment

In the last decades, the development of Autonomous Underwater Vehicles (AUVs) has advanced significantly, focusing on their ability to autonomously navigate and collect data in underwater environments. However, these tasks are often predefined and limited in decision-making capabilities, necessitating human intervention in many industrial operations such as inspection and maintenance, environmental monitoring, disaster response, and port security. Initiatives such as the Metrological Evaluation and Testing of Robots in International Competitions (METRICS), supported by the European Commission, are advancing research in this area by hosting events like the Robotics for Asset Maintenance and Inspection (RAMI) competition. This competition challenges participants to create perception algorithms that can accurately identify a predefined set of targets in underwater imagery. In the present work, a comparison of three different deep learning strategies for detection and classification of artificial objects in marine environments is performed utilizing a custom dataset provided by the NATO Science and Technology Organization Centre for Maritime Research and Experimentation (STO CMRE). In particular, the performance of various Convolutional Neural Networks (CNNs), including Faster Region-based CNN (Faster R-CNN) with VGG19 and ResNet50 backbones, and YOLOv5 architectures were evaluated. The results show that YOLOv5 achieves the highest mean Average Precision (mAP) of 97.2%, even with less computation time. This demonstrates its superior performance in real-time object detection tasks under challenging underwater conditions and scarce data availability.