<![CDATA[Reliable vessel surveillance requires robust computer vision for accurate ship detection. Although YOLO-FAST is computationally efficient, its lightweight architecture has limited capability to extract discriminative features from small objects. To address this limitation, this study proposes an enhanced ship detection framework by integrating a novel Hadamard-Dilated Residual Contextual Transformer (HDR-CoT) module into the YOLO-FAST architecture. The proposed HDR-CoT module integrates Hadamard multiplication with parallel dilated depthwise convolutions. In addition, residual connections enhance feature representation while maintaining low computational overhead. Experimental results demonstrate the effectiveness of the proposed model. It achieves an mAP50 of 0.952 and an mAP50:95 of 0.689 while requiring only 1.73 million parameters and 6.3 GFLOPs. The proposed network achieves a favorable balance between detection performance and efficiency, improving mAP50:95 by 2.99% while reducing the parameters by 24.93%. It also achieves competitive performance compared with other lightweight architectures. CPU-based evaluation further shows that the model operates at 20.54 FPS. This configuration achieves an optimal trade-off between near-real-time inference speed and detection reliability compared with YOLO-FAST for resource-constrained maritime applications.]]>
Copyrights © 2026
