HIFNet: wavelet transform-enhanced UAV object detection in complex conditions
摘要
With deep learning advancements, UAV object detection has great potential in various fields. However, in complex conditions, such as nighttime or fog, small object details are easily obscured by the background, leading to feature loss. To address this, we propose a network based on Wavelet Transform (WT), called the Hierarchical Interactive Fusion Network (HIFNet). First, Wavelet Spatial Frequency Awareness (WSFA) is created to enhance the representation of small objects in complex backgrounds. WT captures edge textures using high-frequency components, while the attention mechanism focuses on target areas. WSFA combines these, using attention weights from high-frequency features to specifically enhance small object expression. Second, Hierarchical Interactive fusion feature pyramid network uses hierarchical interactive fusion technology. Cross-layer fusion enhances shallow and deep features in the backbone, while cross-stage fusion combines shallow details with deep semantics. Then, element-wise addition is used for deep interaction, preserving small details during feature sampling. Finally, the adaptive mixed (A-mixed) loss combines IoU and normalized Wasserstein distance (NWD) losses. It emphasizes small objects by adjusting NWD based on object area and dynamically balances the losses for optimization. Experimental results show HIFNet’s exceptional performance on three datasets, achieving