Pretrained U-Net: in-depth analysis of binary image segmentation in underwater marine environment

Light absorption and scattering, which alter image clarity and colour, present substantial obstacles to segmentation in computer vision, especially in underwater photography. This paper provides a new method for segmenting underwater images using the Keras-UNet architecture. Our proposed method incorporates hybrid loss functions to enhance segmentation accuracy. The hybrid loss functions, combining binary cross-entropy and Jaccard distance, address the unique constraints of underwater imagery by balancing the need for accurate boundary delineation and handling class imbalance. Our method is trained for binary segmentation on the NAUTEC-UWI Real dataset, which includes a diverse array of underwater artefacts and species, such as ruins, statues, wrecks, vertebrates, and invertebrates. The suggested model attains remarkable performance measures, such as an average score of 89.75% for Intersection over Union (IoU) and 94.59% for F-scores. Furthermore, the model successfully recognizes items in photos with good recall (93.38%) and precision (95.83%). On fresh photos, generalization ability is shown to be 93.26% accurate. We verify our method’s efficacy in precisely segmenting underwater imagery by rigorous testing and assessment, advancing the fields of marine biology, oceanography, and underwater robotics in the process.