Enhancing Rotated Object Detection in Remote Sensing With a Parallel Hybrid Attention Mechanism

Accurate detection of rotated objects in remote sensing imagery remains challenging due to densely distributed targets, complex and cluttered backgrounds, and high foreground-background similarity. These factors amplify boundary ambiguity and noise interference, necessitating detection frameworks with enhanced robustness and precision. This study presents a Parallel Hybrid Attention (PHA) module that advances feature extraction for rotated object detection. Unlike conventional sequential attention designs, PHA integrates the Shift Window-Mixed Attention Mechanism (SW-MAM) and Coordinate Attention (CA) within a parallel fusion architecture, effectively preserving complementary spatial-channel information and mitigating feature degradation during transmission. To strengthen multi-scale feature representation, PHA is embedded into a Path Aggregation Feature Pyramid Network (PAFPN), facilitating reliable detection across diverse object scales and rotation angles. Comprehensive experiments on four benchmark datasets, MVTec, HRSC2016, HRSC2016-MS, and DOTA-v1.0, demonstrate the proposed method’s superior performance, achieving mean Average Precision (mAP) scores of 0.920, 0.971, 0.800, and 0.807, respectively. These results confirm PHA’s ability to maintain high detection accuracy in dense and cluttered scenes, establishing it as a scalable and generalizable solution for advanced rotated object detection in remote sensing applications. Our code is available at https://github.com/RVL224/PHA_YOLO2.