Abstract
Detecting small aircraft in remote sensing imagery remains a formidable challenge due to limited pixel representation, complex background interference, and multi-scale variations. Existing detection frameworks often suffer from the loss of high-frequency details during feature extraction and inadequate multi-scale feature fusion, leading to degraded performance on small and densely distributed aircraft targets. In this paper, we propose AFRDet, a novel detection framework built upon RT-DETR that addresses these limitations through two synergistic modules. First, we introduce the Adaptive Frequency Reconstruction Module (AFRM), which decomposes multi-scale features into high-frequency and low-frequency components via learnable filter generators, enabling content-adaptive frequency separation and reconstruction that preserves fine-grained structural details critical for small aircraft localization. Second, we design the Dual-path Gated Fusion Module (DGFM), which employs parallel depthwise separable convolution paths with element-wise gating to selectively aggregate complementary features while suppressing background noise. Furthermore, we use a Scale-aware Normalized Wasserstein Distance (SA-NWD) as a training-time auxiliary localization loss for small aircraft targets, which replaces the conventional IoU-based metric with a Gaussian distribution distance measure to alleviate the sensitivity of bounding box regression to minor positional deviations. Extensive experiments on three benchmarks, including MAR20, UCAS-AOD, and the Jilin-1 satellite aircraft detection dataset, demonstrate that AFRDet achieves state-of-the-art performance. On MAR20, our method attains 85.7% mAP50, surpassing the baseline RT-DETR by 4.2%. On UCAS-AOD, AFRDet achieves 97.3% mAP50, and on the challenging Jilin-1 dataset, it reaches 72.8% mAP50, outperforming all compared methods.