DFDP-Net
Description
Infrared target detection from UAV platforms is important for surveillance and emergency response. However, high-altitude top-down infrared imagery remains challenging because targets are often tiny, weak-textured, and low-contrast, while the background is highly non-stationary and cluttered. To address these issues, we propose the Dynamic Frequency Domain Perception Network (DFDP-Net), which is built on DEIM and introduces frequency-domain cues into key stages of feature encoding. Specifically, we design a Local Feature Enhancement Downsampling Module (LFEDM) to preserve target-related high-frequency details during downsampling, an Adaptive Kernel Transformation Convolution Module (AKTCM) to generate frequency-guided and spatially adaptive filtering responses, and a Channel-Frequency Fusion Attention Module (CFFAM) to enhance target saliency in a joint channel-frequency space.