An Adaptive Network for Arbitrary-Oriented and Small Object Detection in Remote Sensing Images

To address the challenges posed by detection of arbitrary orientation object and small object in remote sensing imagery, this paper proposes an innovative object detection model based on YOLOv8, named YOLOv8-DFS. To improve robustness against arbitrary orientation object, we skillfully design a new model by inserting three SPPELAN modules and a C2f-DCN module into the model. The model not only mines the context of multi-scale and multi-orientation objects through employing weighted fusion of features across different scales, but also enables dynamic adaption of convolution kernel shapes and orientations, thereby significantly enhancing the model’s precision in detecting rotated and inclined targets. To improve the detection performance for small-scale objects, we design IADown and BiConcat modules. The IADown module refines the downsampling strategy, preserving more detailed information and enhancing feature representation for small objects. Concurrently, the BiConcat module facilitates optimized multi-scale feature fusion through a learnable weighted concatenation mechanism, amplifying detailed information necessary of small object recognition. The experimental results on the DIOR-R dataset show that YOLOv8-DFS-s (based on YOLOv8-s) achieves a 2.4% improvement in mAP@0.5 over the baseline, while YOLOv8-DFS-l (based on YOLOv8-l) yields a 1.5% increase. On the HRSC2016 dataset, the results of YOLOv8-DFS-s and YOLOv8-DFS-l show an improvement of 3.1% and 2.0%, respectively. These results validate the efficiency of YOLOv8-DFS in handling both rotational variance and small object detection, demonstrating substantial performance gains across multiple model scales.