SIAM-WFEF: Small target tracking network based on wavelet pooling layer and frequency band enhancement and fusion

TY - GEN

T1 - SIAM-WFEF

T2 - International Conference on Machine Learning and Artificial Intelligence Applications, MLAIA 2025

AU - Meng, Yaomiao

AU - Han, Yuqi

AU - Tang, Linbo

AU - Jin, Qiuyu

AU - Thapa, Sharan

N1 - Publisher Copyright: © 2026 SPIE.

PY - 2026/3/9

Y1 - 2026/3/9

N2 - Unmanned aerial vehicles (UAVs) are increasingly used for target tracking, but long-distance/high-altitude imaging often yields tiny, blurred targets with few pixels. Conventional feature extraction struggles to capture local details and discriminative cues, making targets vulnerable to background interference and reducing tracking accuracy. This paper proposes Siam-WFEF, a Siamese network-based tracker featuring two synergistic modules: Wavelet Pooling Layer (WPL) and Frequency Band Enhancement and Fusion (FBEF). WPL replaces traditional pooling by applying the Discrete Wavelet Transform to decompose feature maps into low- and high-frequency sub-bands, preserving critical details of small targets. FBEF then refines each sub-band and adaptively fuses the enhanced frequency features, jointly suppressing background clutter while amplifying target cues. Extensive experiments on UAV123, DTB70, and VisDrone2019-SOT demonstrate that Siam-WFEF achieves significant performance gains over mainstream trackers, particularly for small targets under challenging conditions.

AB - Unmanned aerial vehicles (UAVs) are increasingly used for target tracking, but long-distance/high-altitude imaging often yields tiny, blurred targets with few pixels. Conventional feature extraction struggles to capture local details and discriminative cues, making targets vulnerable to background interference and reducing tracking accuracy. This paper proposes Siam-WFEF, a Siamese network-based tracker featuring two synergistic modules: Wavelet Pooling Layer (WPL) and Frequency Band Enhancement and Fusion (FBEF). WPL replaces traditional pooling by applying the Discrete Wavelet Transform to decompose feature maps into low- and high-frequency sub-bands, preserving critical details of small targets. FBEF then refines each sub-band and adaptively fuses the enhanced frequency features, jointly suppressing background clutter while amplifying target cues. Extensive experiments on UAV123, DTB70, and VisDrone2019-SOT demonstrate that Siam-WFEF achieves significant performance gains over mainstream trackers, particularly for small targets under challenging conditions.

KW - Siamese network

KW - UAV vision

KW - frequency fusion

KW - small target tracking

KW - wavelet pooling

UR - https://www.scopus.com/pages/publications/105034715607

U2 - 10.1117/12.3110411

DO - 10.1117/12.3110411

M3 - Conference contribution

AN - SCOPUS:105034715607

T3 - Proceedings of SPIE - The International Society for Optical Engineering

BT - International Conference on Machine Learning and Artificial Intelligence Applications, MLAIA 2025

A2 - Zhou, Jianhua

PB - SPIE

Y2 - 12 December 2025 through 14 December 2025

ER -