IRKAN: Wavelet-Based Kolmogorov–Arnold Networks for Infrared Small Target Detection

Infrared small target detection is fundamentally challenged by the poor generalization of deep learning models across diverse sensors and imaging resolutions. To address this limitation, a novel framework named IRKAN is proposed, which reimagines the network’s core feature extraction mechanism. The proposed approach moves beyond conventional architectures by adopting the Kolmogorov-Arnold Network paradigm, replacing fixed, node-based activations with learnable functions on the network edges. This design endows the model with the intrinsic capability to learn optimal, data-driven transformations for complex infrared signals. To tailor this adaptive framework for the specific physics of infrared small target detection, wavelet basis functions are employed to parameterize these activations. This innovation infuses a powerful signal-processing prior into the network, leveraging the time-frequency localization properties of wavelets to function as a learnable matched filter, adept at isolating the faint, high-frequency signatures of targets from background clutter. By strategically embedding these specialized wavelet-parameterized Kolmogorov-Arnold Network modules into a U-Net backbone, IRKAN achieves strong performance on the public WideIRSTD-Full dataset. Evaluated over ten independent runs, it attains a mean normalized Intersection over Union of 68.89% and a low False Alarm Rate of 8.33 × 10−6, while maintaining computational efficiency (3.96M parameters). The synergy between the architectural flexibility of Kolmogorov-Arnold Networks and the signal-aware inductive bias of wavelets provides a robust solution to the challenges of modern infrared small target detection. Zero-shot evaluations on the unseen NUDT-SIRSTD (512 × 512) and IRSTD-1K (256 × 256) datasets indicate that IRKAN possesses strong cross-resolution generalization capabilities, particularly under resolution shifts, without requiring fine-tuning.