Integrating Image Enhancement, Classification, and Segmentation for Supporting Automated Underwater Infrastructure Inspection and Maintenance

As climate change accelerates, there is a pressing need to safely and cost-effectively inspect and maintain underwater infrastructure. Recent advancements in computer vision and robotics open unprecedented opportunities for automated underwater inspection and maintenance that is safer and more cost-effective. However, underwater automation faces challenges due to severe visibility issues and color distortion of underwater images, making it challenging to locate structures for subsequent operations. To address these challenges, this paper proposes a new framework to locate structures in underwater images. The proposed framework integrates image enhancement, classification, and segmentation for localization, including the Deep WaveNet model for image visibility enhancement, the MLP-Mixer model for retrieving images containing structures, and ResNet + PSPNet for image segmentation to locate target structures. The preliminary results from ablation analysis show that integrating the three models, compared to the alternatives without one or two models, achieved the highest intersection of over 0.810.