PSWSA-RetinaNet: Polarized Shifted-Window Self-Attention Enhanced RetinaNet for Underwater Object Detection in Complex Marine Environments

Abstract

Underwater Object Detection (UOD) is important for resource exploration, marine conservation, and scientific research. However, the occurrence of closely packed targets becomes an important problem due to greater degree of similarity in geomorphology, spectral and textural characteristics which mostly resulted in inaccurate classification. This spatial overlap mostly reduces the detection performance and impacts bounding box merging, which failed to separate individual targets. Hence, this research proposes the Polarized Shifted-Window Self-Attention-based RetinaNet (PSWSA-RetinaNet) approach, that involves ResNet50 and a Feature Pyramid Network (FPN) for accurate UOD. The proposed approach integrated the PSWSA with RetinaNet for enhancing the discriminative feature representation and an SWSA for capturing the long-range contextual dependencies. Through incorporating these components, the proposed approach improves sensitivity to small object detection whereas handling the important features which are cooperated through aquatic minimization. The experimental findings demonstrate that the proposed PSWSA-RetinaNet approach attain the precision of 91.40% and 72.70% on the Detecting Underwater Objects (DUO) and TrashCan datasets, separately. The proposed approach illustrates enhanced performance when compared with existing configurations under a controlled estimation component. These results confirm that PSWSA-RetinaNet enhances detection robustness over different benchmark datasets under the reported estimation component.