FSCR-Net is proposed, a frequency-sensitive contour reconstruction network outperforming state-of-the-art approaches in both detection accuracy and contour representation quality, outperforming state-of-the-art approaches in both detection accuracy and contour representation quality.
Abstract
Rotated object detection in remote sensing imagery is a core task of intelligent interpretation, used in urban management, resource surveys, and national defense. Remote sensing targets have large scale gaps, arbitrary orientations, and rotational symmetry, demanding robust feature extraction and shape description. Existing methods have two limitations: first, they cannot distinguish tiny targets’ valid high-frequency features from background clutter. Global frequency-domain filtering impairs large-object features, raising small-target miss rates. Second, rotated bounding boxes fail to depict irregular ground-object contours, and annotation starting points degrade training stability. To address these, we propose FSCR-Net, a frequency-sensitive contour reconstruction network. Leveraging Fourier series’ periodic symmetry for robust parametric contour representation, it improves detection and contour quality via three modules. First, a frequency-smoothness-aware feature fusion module decouples foreground/background features via soft masks and applies scale-adaptive frequency-domain filtering on background branches, suppressing clutter and enhancing tiny target discriminability without impairing large objects. Second, a differentiated contour representation strategy converts contours into parametric Fourier series, achieving unified dimensional normalization for multi-scale targets and compact shape encoding for detection regression. Third, a Fourier coefficient reconstruction detection head adopts a rolling optimization matching loss to eliminate annotation starting point interference and reconstructs contours via inverse Fourier transform. The proposed method achieves 76.52% mAP@0.5 on DOTA-v1.5, 92.15% mAP on HRSC2016, and 98.71% mAP on UCAS-AOD, outperforming state-of-the-art approaches in both detection accuracy and contour representation quality.
Remote sensing objects are characterized by arbitrary orientations, complex backgrounds, and high aspect ratios (ARs). Most existing methods rely on explicit discrete convolutions to construct feature networks, which suffer from limitations such as insufficient rotation-invariant representation, background noise introd...
Yan Dong, Xiang-Yun Li, Guang-Shuai Gao et al.· IEEE Transactions on Geoscie...· 0 citations
Remote sensing small-object detection remains challenging: weak target details are attenuated by repeated downsampling, sparse object responses are difficult to separate from background clutter, and objects vary widely in scale and aspect ratio. We propose FD-YOLO, a compact detector with three complementary components...
Qi-Guang Zhu, Yan Zhu, Xing-Yong Lv et al.· Measurement science and tech...· 0 citations
Remote sensing object detection plays a crucial role in national defense and security applications. Although anchor-free methods such as CenterNet++ achieve competitive recall and low false-positive rates, they remain limited by restricted global receptive fields and insufficient exploitation of geometric contour infor...
Remote sensing oriented object detection faces the challenge of maintaining consistent rotational geometric information throughout feature representation, cross-scale propagation, and bounding-box localization. To address this issue, this paper proposes a direction–scale collaborative optimization method for oriented o...
Bin Zhang, Wei-Guo Pan, Bing-Xin Xu et al.· Remote Sensing· 0 citations
COF-YOLO consistently outperforms the baseline YOLOv12 and several mainstream lightweight detectors, especially under small-target and occluded conditions, while maintaining practical inference efficiency.
Bo-Yu Liu, Xu-Yang Li, Rui Cao et al.· IEEE Access· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.