Recent developments in deep-sea exploration, environmental surveillance, and ocean research, accurate segmentation of underwater images is essential. This study pursues this goal by exploring underwater image segmentation from a deep learning perspective. It specifically looks at how well the Attention Residual UNet architecture, a more sophisticated version of the U-Net, works with the focal loss technique to achieve accuracy in this crucial task. Using attention mechanisms and residual connections, the Attention Residual UNet architecture, based on the U-Net framework, can capture fine details while maintaining contextual coherence. Model loss and accuracy measures are considered as this study carefully assesses the architecture’s performance. Notably, the model exhibits outstanding accuracy, obtaining 97.23% and 92.76% accuracy on the training and validation sets, respectively. The Jaccard coefficient further demonstrates the model’s efficiency, which measures the intersection between predicted and actual segments and has coefficients for the model and validation sets of 74.31% and 66.51% , respectively. The Mean Intersection over Union (MIoU) statistic, which boasts a value 91.24%, validates the model’s superiority.