Adaptive Fusion Network of Deep Frequency-Domain Features for Infrared Small Target Detection

Infrared small target detection holds significant application value in defense security, intelligent surveillance, and other fields. However, due to the low contrast of small targets in infrared imagery, complex and variable backgrounds, and low signal-to-noise ratios, it remains a challenging task in computer vision. Existing methods generally employ deep learning techniques to improve detection accuracy, yet they often suffer from high computational complexity, inefficient inference, and insufficient sensitivity to small targets, failing to meet the dual requirements of real-time performance and precision in practical applications. To address these issues, this paper proposes an infrared small target detection algorithm based on deep frequency domain feature fusion, effectively balancing efficiency and accuracy. First, we employ the computationally efficient MobileNetV3 as the backbone network, significantly reducing model parameters and computational cost, establishing a foundation for deployment in resource-constrained environments. Second, to address the critical issue of small target feature extraction, we propose a Multi-Branch Frequency-Aware Module (MBFAM). This module expands the receptive field through convolution operations with different dilation rates and decomposes frequency domain information via Haar Discrete Wavelet Transform (HDWT), effectively capturing weak features of small targets and enhancing the model's perception of small-scale objects. Finally, in the multi-scale feature fusion stage, we innovatively adopt a dynamic fusion strategy based on Kolmogorov–Arnold Networks (KAN) to replace traditional Multi-Layer Perceptron, adaptively allocating weights to different feature layers, highlighting the contribution of key features while suppressing interference from redundant information. The experimental results on multiple datasets show that the proposed method has more accurate detection performance and faster running speed compared to the existing state-of-the-art methods.