PQGNet: Perceptual Query Guided Network for Infrared Small Target Detection

Infrared small target detection (IRSTD) holds critical importance for military security applications. Although U-shaped architectures have improved baseline performance, existing methods still suffer from two key limitations: 1) Insufficient spatial perception for tiny targets leads to target location loss.; 2) Edge degradation and semantic ambiguity in deep feature reconstruction. To address these challenges, we propose PQGNet with the following contributions: To enhance capability of spatial perception and improve feature fusion guidance, we introduce the Perceptual Query Supervision Mechanism (PQSM), which utilizes perceptual loss to constrain spatial feature learning of each encoder layer. The Perceptual Feature Construction Module (PFCM) constructs enhanced perceptual features to preserve target localization information, while the Perceptual Query Guidance Module (PQGM) adopts cross-attention to guide global and regional feature queries through skip connections, optimizing target feature extraction. To mitigate reconstruction degradation and semantic ambiguity, distinct from existing wavelet-based approaches that simply substitute pooling layers, we design a Max pooling-Wavelet Hybrid Layer (MWHL) and High-frequency Enhancement Wavelet Layer (HEWL) that exploit discrete wavelet transform properties to enhance deep semantic representations using shallow high-frequency details. Comprehensive experiments on NUDT-SIRST, NUAA-SIRST, and IRSTD-1K datasets demonstrate that PQGNet significantly surpasses state-of-the-art methods in detection performance, while maintaining a competitive balance between computational complexity and accuracy. Our code will be made public at https://github.com/PepperCS/PQGNet.