.png){kind=icon}
.png){kind=icon}
Return
ETD-Det: Oriented Object Detection With Spectral Diffusion Encoding and Extended-Gaussian Decoding
.png){kind=icon}
PRE
AI
DOI：10.1109/TGRS.2026.3663195.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Background noise contamination in encoder features and insufficient directional modeling on the decoder side remain central challenges in arbitrary-oriented object detection (AOOD). To address these issues, an encode–transmit–decode detector (ETD-Det) is presented, co-optimized across encoding, transmission, and decoding. On the encoding side, a spectral diffusion encoder (LapSpecEnc) is proposed. Frequency-domain diffusion is performed by LapSpecEnc to suppress high-frequency background noise and to maintain coherent, orientation-consistent features. On the decoding side, an extended-Gaussian distribution (EGD) together with a Kullback–Leibler (KL) divergence (EGKLD)-based loss is introduced. This design provides direction-sensitive geometric representation and mitigates the angle-gradient degradation observed in standard Gaussian formulations, particularly for nearly square targets. Extensive evaluation has been conducted, demonstrating strong effectiveness and generalization: ETD-Det achieves 90.61% mAP on HRSC2016, 90.37% mAP on UCAS-AOD, 81.54% mAP on DOTA-v1.0, and an F-measure of 82.33% on ICDAR2015. These results surpass existing methods under comparable settings and support ETD-Det as a robust and accurate solution for high-precision AOOD.
Keywords:
Angle gradient
arbitrary-oriented object detection (AOOD)
extended-Gaussian distribution (EGD)
Kullback–Leibler (KL) divergence
Laplacian spectral backbone
remote sensing
scale invariance
spectral diffusion
Journal
IF:
8.6
Papers:
2.1W
Citations:
10.7W