2026· IEEE Transactions on Geoscience and Remote Sensing· Vol 64, pp. 4707214-4707214· 0 citations· 46 references
Abstract
Small object detection in remote sensing imagery remains challenging due to the extremely limited pixel footprint of targets and the severe loss of fine-grained spatial details caused by multistage downsampling. To address these issues, we propose WE-DETR, a wavelet-enhanced multiscale feature compensation detection transformer. WE-DETR establishes a remote sensing-oriented progressive high-frequency (HF) feature compensation mechanism, which aims to preserve, enhance, fuse, and propagate discriminative structural cues throughout the detection pipeline. In particular, wavelet decomposition is introduced at the early backbone stage to preserve fine-grained details before downsampling-induced degradation occurs. A low-frequency-guided structural enhancement strategy is further developed to selectively strengthen informative HF components. The enhanced frequency-domain representations are then adaptively integrated with spatial-domain semantic features to improve detail preservation while suppressing clutter-induced responses. Finally, the compensated shallow details are propagated into multiscale detection features using detail-preserving downsampling and efficient contextual enhancement. Extensive experiments on multiple remote sensing benchmarks, including optical, infrared, and synthetic aperture radar (SAR) datasets, demonstrate that WE-DETR achieves competitive performance compared with state-of-the-art detectors while maintaining a favorable balance between accuracy and computational cost. These results validate the effectiveness of the proposed wavelet-enhanced multiscale feature compensation framework across representative remote sensing small object detection tasks. The code is available at https://github.com/jingmingliang/WE-DETR
The feature mapping stage is reconstructed to incorporate physical edge priors and frequency-domain attention-based denoising, and a Spatial Attention Wavelet Transform Feature Decomposer (SA-WTFD) is introduced in the frequency branch to adaptively suppress high-frequency background clutter and enhance discriminative...
Qi-Yuan Zhang, Jian-Shun Liu· Italian National Conference...· 1 citation
Small object detection in remote sensing (RS) imagery remains fundamentally challenging due to severe information degradation caused by limited spatial resolution and complex background interference. In deep neural networks, such degradation is further exacerbated by irreversible information loss during conventional do...
Ying Gao, Zongshuai Zhang, Zheng-Yu Zhu et al.· IEEE Transactions on Geoscie...· 0 citations
Due to the severe scale variation of targets in remote sensing images, the dense distribution of objects, and the fact that many small targets occupy only a very limited number of pixels, existing detection methods are prone to losing shallow details and suffering from insufficient low-level semantic representation dur...
Fa-Quan Song, Wu Le, Ming Lv et al.· IEEE Transactions on Geoscie...· 1 citation
This work proposes a novel architecture that leverages Wavelet Transform Fusion (WTF) and Context-Guided Downsampling (CGD) in the real-time detection transformer (RT-DETR) for small object detection and demonstrates that CGWT-DETR outperforms the RT-DETR baseline in both detection accuracy and computational efficiency...
Shahzad Hussain, Iqra Mumtaz, Usman Ahmad et al.· Remote Sensing· 1 citation
A deep learning framework is introduced, which leverages a hierarchical frequency-domain decomposition to simultaneously preserve semantic context and fine-grained textures to simultaneously preserve semantic context and fine-grained textures.
The results demonstrate that MWAE-YOLO can effectively enhance frequency-aware small-object feature representation, suppress background interference, and improve scale-sensitive localization accuracy in complex remote sensing scenes.