Learning DCT Subband using Kolmogorov-Arnold Network for Infrared Small Target Detection

Deep learning-based infrared small target detection (IRSTD) methods typically exploit spatial domain cues to infer dim and weak infrared targets. However, relying solely on spatial domain information is sub-optimal due to the lack of structure and texture details of the target. Alternatively, we learn frequency priors to enhance targets in frequency domain, and propose a Frequency Domain Enhanced Network (FDENet). It contains a frequency domain discrete cosine transform enhancement (FDE) module with KAN-inspired DCT learning mechanism. Specifically, the FDE module uses multiple univariate nonlinear functions to combine continuous multivariate functions to directly map frequency-aware cues without spatial features. It then aligns the frequency-aware cues and with the spatial features to capture dimly weak targets. Generalization experiments and comparative studies on two real-world infrared single-frame image datasets fully demonstrate the effectiveness of our method.