FCD-YOLO: An Oriented Object Detector Based on Feature Extraction and Context Enhancement

Persistent challenges in remote-sensing detection include pronounced variations in object scale, cluttered backgrounds with strong interference, and the prevalence of numerous, densely distributed small objects. Under the oriented-object detection paradigm, the additional orientation dimension renders the task even more demanding. To address this challenge, a feature extraction and context-enhanced YOLO variant (FCD-YOLO) was devised. Context Aggregation Pinwheel-shaped Convolution (CAPConv) incorporates an attention mechanism with an asymmetric padding design, enabling comprehensive context perception, significantly expanding the receptive field, and enhancing the detection accuracy of small, directional targets. Rotating object detection necessitates handling numerous direction-dependent features. The Adaptive Rotation Feature Extraction (ARFE) module employs adaptive rotation convolutional kernels to match the rotation angle of the target, thereby enabling comprehensive extraction of directional features. Global-Local Feature Aggregation (GLFA) simultaneously processes global and local features, enabling the model to comprehensively understand contextual information. Dilation convolutions and multi-scale pooling operations extract features across multiple scales, mitigating the impact of varying object sizes on detection. Finally, experiments conducted on the publicly available large-scale rotated object detection dataset DIOR-R demonstrate that FCD-YOLO achieves mAP@0.5 of 70.5%, representing 1.9% improvement over the baseline model.