Enhancing Aerial Ship Segmentation: Attention-Based U-Net Optimization with Reduced Resolution

Ship detection plays a crucial role in maritime surveillance, ecological factors, maritime security, and safety application. Thus, helps in detecting threats, illegal activities and standardizes a lot of motion in sea. But it is challenging due to complex ship boundaries and occlusions. To address these issues, attention and residual mechanisms are integrated into the U-Net architecture, focusing on relevant regions, and improving feature representation. Several U-Net variants are proposed and evaluated on a diverse dataset of aerial ship images. The experiment is presented on satellite image datasets where various encoder, decoder and loss function are evaluated on UNet models. The objectives of this study are to evaluate and compare the dice coefficient and the Jaccard coefficient, as well as apply variations of U-Net models on lower resolution images. The best-performing model is Residual Attention U-Net (RAUNet), which takes less time and loses 3% on dice and Jaccard values from 256X256 to 128X128 resolution.