CNet: A Novel Seabed Coral Reef Image Segmentation Approach Based on Deep Learning

Hanqi Zhang, Ming Li, Jiageng Zhong, Jiangying Qin; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2024, pp. 767-775

Achieving underwater coral seabed image segmentation involves dividing an image into meaningful regions or segments, which, in this case, could represent different types of corals, substrate, or other features in the underwater habitat. We introduce an innovative network architecture, CNet, designed for the segmentation of coral seabed images. This architecture incorporates a three-branch parallel encoder structure, employing an RGB encoder based on the ResNet block, a Depth encoder based on the VGG block, and a ShapeConv block-based Fusion encoder. The study conducts comprehensive performance comparisons and ablation experiments to evaluate the efficacy of CNet in comparison to state-of-the-art (SOTA) methods. The results demonstrate an impressive mIoU of 81.83% on the coral dataset, with the IoU of the minority class, Acropora, reaching 73.61%. This is of crucial significance in the fields of marine biology and environmental monitoring, playing a pivotal role in the comprehensive understanding of coral reef ecosystems. By automatically and accurately identifying different coral classes, scientists can gain insights into threatened corals and their growth in different environments, providing crucial data for developing targeted conservation plans to promote coral recovery.