ABSTRACT:

Rotated object detection in remote sensing images is critical for applications such as maritime monitoring, airport surveillance, and aerial reconnaissance. Traditional detection approaches often struggle with accurately identifying objects that have arbitrary orientations, leading to reduced detection performance. This research presents a novel approach by integrating a parallel hybrid attention mechanism into two deep learning frameworks: YOLO26_PHAM and YOLO11_SDA. YOLO26_PHAM leverages a hybrid attention module combining CBAM and SE blocks to capture channel and spatial dependencies, enhancing feature representation in backbone layers. YOLO11_SDA employs Spiral Depthwise Attention to extract multi-scale contextual information efficiently, improving detection of small and rotated objects. The models are trained on a merged dataset containing ships and airplanes, with images labeled in YOLO format. The system is deployed with a Flask-based backend and a web interface developed using HTML, CSS, and JavaScript, enabling efficient image processing and prediction display. Experimental results show improved localization accuracy, higher intersection-over-union scores, and better generalization across orientations. The integration of hybrid attention mechanisms demonstrates a significant enhancement in detecting rotated objects compared to baseline YOLO models, providing an effective solution for remote sensing detection tasks.

Keywords: Rotated Object Detection, Remote Sensing, YOLO26_PHAM, YOLO11_SDA, CBAM, SE Block, Spiral Depthwise Attention, Hybrid Attention, Ship Detection, Airplane Detection