MST-DETR: A multi-scale enhanced tiny object detection framework

Lingling Li

doi:10.1007/s11760-025-05074-8

Return

MST-DETR: A multi-scale enhanced tiny object detection framework

2026-01-01

PRE

DOI：10.1007/s11760-025-05074-8

Original

Original request for help

Save

Abstract

En 中文

Tiny object detection is a critical challenge in computer vision with significant applications in aerial surveillance and emergency response systems. Existing methods suffer from three persistent limitations: indistinct feature representations for small-scale targets, severe background interference, and structural detail degradation during multi-scale processing. To address these issues, we propose MST-DETR, a multi-module collaborative framework with three key innovations. First, the Adaptive Multi-Scale Saliency Enhancement (AMSE) module dynamically adjusts feature fusion weights through spatial attention, enhancing discriminative object characteristics while maintaining computational efficiency. Second, the Efficient Upsampling Feature Reorganization Module (EUFRM) improves feature alignment via progressive upsampling and optimized channel interactions, boosting classification robustness in complex environments. Third, the Dynamic Partial Downsampling (DPD) module integrates deformable operations with wavelet-based processing to minimize structural information loss during feature extraction. Extensive experimental results demonstrate that MST-DETR achieves a 2.5%-7.2% improvement in mAP50val\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$mAP_{50}<^>{val}$$\end{document} compared with RT-DETR-R18 on the SIMD and VisDrone2019 datasets. Meanwhile, it significantly reduces false positives and missed detections, further validating the effectiveness and robustness of the proposed method in tiny object detection tasks.

Keywords:

Tiny Object Detection

Multi-Scale

DETR