ICFIFormer: Illumination Compensation and Asymmetric Cross-Band Frequency Interaction for Low-Light Image Enhancement

Abstract

Low-light image enhancement remains challenging due to nonuniform illumination, color distortion, and noise-induced detail degradation. Existing methods often emphasize either illumination modeling or frequency-domain restoration, which limits their ability to jointly perform illumination compensation and detail restoration. We propose the Illumination Compensation and Frequency Interaction Former (ICFIFormer), a one-stage compensation–restoration framework for low-light image enhancement. The Illumination Compensation Subnetwork employs quaternion convolution to jointly model the color prior and RGB channels, and predicts a light-up map to compensate for nonuniform illumination in the input image. In the restoration stage, Frequency-Interactive Cross-Attention (FICA) constructs queries from the low-frequency subband and keys and values from the concatenated high-frequency subbands, thereby realizing spatially weighted asymmetric cross-band transposed channel attention. The Illumination-Compensation Guided Feed-Forward Network (ICFFN) incorporates the light-up map into feature updates. Experiments on five benchmark datasets demonstrate that ICFIFormer achieves competitive performance against recent low-light image enhancement methods.