Enhancing Underwater Image Segmentation: A Semantic Approach to Segment Objects in Challenging Aquatic Environment

Abstract

Aquatic image segmentation is complex due to the inherent complexity of light attenuation, color shifts, and vision limits in underwater settings. In order to improve underwater image segmentation and the delineation of objects in such conditions, this research study suggests a novel semantic technique. The suggested approach uses cutting-edge deep learning methods and a U-Net architecture with DenseNet-201 at its core. The dense connectivity patterns of DenseNet-201 make it possible to extract useful features, increasing gradient flow and feature reuse across the network. The model uses a combination of Dice Loss and Categorical Focal Loss is the loss function to enhance segmentation performance. It utilizes evaluation metrics like the IoU Score and F-Score, which give detailed insights into the model's performance on underwater picture segmentation tasks. According to the evaluation metrics, the "DenseNet201" architecture is the most effective one for underwater picture segmentation. On the test dataset, it received the highest Mean IoU Score (79.77/%) and F1 Score (84.85)%. Furthermore, it obtained competitive results for each class-wise IoU score, demonstrating strong segmentation performance for different object categories.