YOLO-SAD: Enhancing Small Aircraft Detection With Multi-Scale Context and Improved Gradient Flow

Abstract

Small object detection, especially for fast-moving targets like small aircrafts, faces significant challenges due to limited texture information, motion blur, and complex background environments. In this study, we propose YOLO-SAD, a novel object detection framework designed to enhance the accuracy and efficiency of small aircraft detection through innovative mechanisms. YOLO-SAD comprises three core modules: The Cross-Scale Attention Kernel Aggregation Module enhances small object feature representation by leveraging complementary channel information through a cross-attention mechanism, improving feature fusion and context enhancement; The Feature Highway Aggregation Structure seamlessly integrates features from different network levels while preserving critical information and improving gradient flow behavior in the shallow layers by utilizing auxiliary training branches during training; The Mixed Spectral Feature Compression Module combines multiple downsampling strategies, introduces multi-scale contextual factors, and retains more original details, effectively reducing information loss. Experimental results on the Det-Fly dataset show that YOLO-SAD achieves a 3.9% improvement in AP and a 3.0% increase in recall compared to YOLOv11L. Additionally, generalization performance on the VisDrone2019-DET dataset indicates a 1.8% improvement in small object AP and a 14% reduction in GFLOPS. These results emphasize the robustness and efficiency of YOLO-SAD, establishing a new performance benchmark for small object detection in complex real-world scenarios.