Infrared Target Detection in UAV Imagery with Wavelet-Based Multi-Level Feature Fusion Network

With the development of infrared imaging technology, infrared target detection (IRTD) has become a key area in computer vision. Due to the low contrast, high noise, and weak distinction between target and background radiation in infrared images, traditional image processing methods face significant challenges. Although deep learning, particularly Convolutional Neural Networks (CNNs), has improved infrared small target detection, issues like scale variations, background interference, and weak image textures continue to hinder performance. In this work, we propose an infrared target detection method by integrating a wavelet-based high-frequency attention mechanism (WHA) with a multi-scale feature fusion module (WSAF) to enhance feature extraction and robustness under low contrast and interference. In order to further improve localization accuracy, we design a hybrid regression loss that combines Location Sensitive Loss, Wasserstein Distance, and Complete IoU, ensuring precise alignment in terms of geometry, spatial structure, and fine-grained localization. Experimental results on the Drone Vehicle and HIT-UAV datasets demonstrate the effectiveness of the proposed method. On the Drone Vehicle dataset, the proposed method achieves 82.5% mAP0.5, while on the HIT-UAV dataset, it reaches 95.4% mAP0.5, setting a new state-of-the-art. These results confirm that the proposed method improves accuracy and robustness in infrared target detection, especially in challenging conditions such as small-scale targets and weak textures.