HiFi-Deblur: High-Frequency Intense Image Deblurring with Frequency-Decoupled U-Net and Discrete Wavelet Transform

Jeewoo Lim, Cheolhee Yu, Jin Tae Kwak; Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) Workshops, 2026, pp. 984-993

Image blur, often induced by camera shake or object motion, remains a significant challenge in computational photography, leading to the degradation of critical image details. To address this issue, we propose HiFi-Deblur, a U-Net-based framework for high-frequency intense image deblurring. Our method leverages discrete wavelet transform to decompose the blurred image into low- and high-frequency components, enabling frequency-aware processing. In the encoder, a joint architecture combining Transformer layers and discrete wavelet transform captures long-range dependencies and extracts hierarchical highfrequency features by directly propagating high-frequency components across encoder stages. In the bridge, alignment and refinement modules enhance low-frequency features and correct spatial mis-alignments. In the decoder, inverse wavelet transform reconstructs refined and enhanced low-frequency features and high-frequency features, and Transformer layers are used to further refine the fused features, enabling the decoder to effectively recover sharp, detailed, and spatially coherent outputs. Extensive experiments on multiple benchmark datasets, including GoPro, HIDE, RealBlur, and DDPD, demonstrate that HiFi-Deblur achieves state-of-the-art performance. This comprehensive validation underscores the superior performance of our approach in addressing a wide spectrum of real-world blur types.