Advancing underwater binary image segmentation with PDBU-Net: A progressive approach to feature extraction and accuracy improvement

Abstract

Precise binary image segmentation in underwater environments is crucial for marine biology and underwater robotics applications, as fluctuations in lighting, water conditions, and object appearance provide substantial difficulties. To tackle these problems, this research presents PDBU-Net, an innovative deep-learning framework specifically created to improve the extraction of features and the accuracy of segmentation for binary images captured underwater. PDBU-Net incorporates a progressive feature extraction methodology into a strong deep neural network structure, resulting in significant enhancements in segmentation performance compared to current methods. PDBU-Net was tested extensively and achieved an average Intersection over Union (IoU) of 90.39% and an F-score of 94.94%. The model achieved precision and recall rates of 94.55% and 95.39% respectively, along with an overall accuracy of 95.66%. The results validate the efficacy of PDBU-Net in precisely detecting objects in a wide range of underwater photos, showcasing its successful implementation in real underwater analysis and robotics.