WHCD-DETR: Frequency-aware cross-scale diffusion architecture for end-to-end UAV small-object detection

Huaxin Cai

doi:10.1016/j.neucom.2026.134687

WHCD-DETR: Frequency-aware cross-scale diffusion architecture for end-to-end UAV small-object detection

2026-08-05

0

PRE

AI

H

Huaxin Cai

X

Xichen Li

Y

Yingying Li

Y

Yifan Luo

D

Dongfeng Liu *

DOI：10.1016/j.neucom.2026.134687

Original

Original request for help

Save

Abstract

En 中文

• An end-to-end detection framework, WHCD-DETR, is proposed for UAV aerial images with dense small objects and complex backgrounds. • WAA-Net integrates wavelet-domain representations with adaptive feature aggregation to enhance high-frequency details and suppress redundant background responses. • HRFFT adopts a channel-decoupled branched architecture to complement global semantic modeling with local structural refinement under controlled complexity. • CSFID introduces cross-scale interaction and feature-space propagation with dynamic upsampling to improve multi-resolution feature alignment. • On VisDrone2019, WHCD-DETR-S achieves an AP50 of 47.8%, improving by 3.4 points over DFINE-S; WHCD-DETR-N achieves an AP50 of 40.4%, improving by 6.7 points over DFINE-N.

WHCD-DETR: Frequency-aware cross-scale diffusion architecture for end-to-end UAV small-object detection

Abstract

Journal

Organization

Cited Papers

Citing Papers