12 September 2026 in SSRN Electronic Journal
Image restoration tasks such as desnowing and dehazing remain challenging due tocomplex degradations that affect both low- and high-frequency components. ExistingCNN-based methods often rely on Layer Normalization, which enforces fixed statisticaldistributions and can oversmooth fine-grained details, limiting the recovery ofcritical high-frequency information. To address this limitation, we propose a unifiedframework, termed NAFXT, which integrates Dynamic Tanh (DyT) normalization withfrequency-aware modules. Unlike conventional normalization, DyT employs learnable,data-dependent scaling that, when combined with standard convolutional operations,preserves local feature variations and avoids oversmoothing, thereby maintaining essentialhigh-frequency details. This enables the generation of detailed feature mapsthat are essential for subsequent frequency processing. Specifically, the NAFXTBlockacts as a preprocessing stage, while the Multi-Branch Dynamic Selective Frequency(MDSF) and Multi-Branch Compact Selective Frequency (MCSF) modules captureglobal frequency information. This synergistic design facilitates effective local andglobal feature extraction. Experimental results demonstrate that the proposed methodachieves superior performance on both desnowing and dehazing benchmarks, reachingup to 41.25 dB PSNR while training time converges faster than Layer Normalization basedcounterparts. However, the framework has not yet been validated on a broaderrange of restoration tasks, which we leave for future work.