Multi-Head Self Attention for Enhanced Object Detection in the Maritime Domain
Walid Messaoud, Rim Trabelsi, Adnane Cabani, Fatma Abdelkefi · 2023
Maritime object detection is crucial for improving the effectiveness of unmanned surface vehicles (USVs), enhancing situational awareness at sea, and strengthening coastal surveillance. Visual understanding is a promising research thematic that can improve the accuracy of ship detection in maritime environment. However, due to the particularities of the maritime environment and the lack of datasets, this field remains limited.In this paper, we propose using Multi-Head Self-Attention (MHSA) to enhance maritime object detection and address the challenges faced. In addition, we conduct a comparative analysis of publicly available maritime datasets to evaluate the performance of different models. Our findings demonstrate that the DETR (DEtection TRansformer), AACN (Attention Augmented Convolutions Networks), and BoTNet (Bottleneck Transformers) models, based on MHSA, exhibit superior detection accuracy and efficiency compared to other existing methods in the maritime domain. This suggests that MHSA is a highly effective approach for improving maritime object detection. Our analysis of the results also provides insights into potential avenues for future research on attention mechanisms in the maritime domain.