DWDN: Deep Wiener Deconvolution Network for Non-Blind Image Deblurring

TL;DR

This paper addresses non-blind image deblurring by integrating an explicit Wiener deconvolution step into a deep feature space, followed by a multi-scale cascaded refinement network trained end-to-end. The feature-based Wiener deconvolution leverages learned feature mappings F_i to compute latent x_hat via G_i, constrained by a per-feature Wiener filter in the frequency domain, improving artifact suppression and detail preservation. The approach, termed DWDN (and its enhanced DWDN+), combines feature extraction, Wiener deconvolution, and cascaded refinement across scales, achieving state-of-the-art results across simulated Gaussian noise, saturation, JPEG artifacts, and real blur, while remaining robust to noise levels and kernel inaccuracies. The work demonstrates that deconvolving in a deep feature space, together with multi-scale refinement, yields clearer restorations and practical improvements for real-world deblurring tasks, with competitive runtime. $y = x * k + n$ models the degradation, and the proposed framework effectively handles variations in $n$ and $k$ through end-to-end learning and adaptive feature-based deconvolution.

Abstract

We present a simple and effective approach for non-blind image deblurring, combining classical techniques and deep learning. In contrast to existing methods that deblur the image directly in the standard image space, we propose to perform an explicit deconvolution process in a feature space by integrating a classical Wiener deconvolution framework with learned deep features. A multi-scale cascaded feature refinement module then predicts the deblurred image from the deconvolved deep features, progressively recovering detail and small-scale structures. The proposed model is trained in an end-to-end manner and evaluated on scenarios with simulated Gaussian noise, saturated pixels, or JPEG compression artifacts as well as real-world images. Moreover, we present detailed analyses of the benefit of the feature-based Wiener deconvolution and of the multi-scale cascaded feature refinement as well as the robustness of the proposed approach. Our extensive experimental results show that the proposed deep Wiener deconvolution network facilitates deblurred results with visibly fewer artifacts and quantitatively outperforms state-of-the-art non-blind image deblurring methods by a wide margin.

Figures (13)

• Figure 1: Deblurring results on a real blurry image from XuECCV2010. A recent blind image deblurring method ZhangCVPR2019, based on end-to-end trainable networks, does not effectively estimate a clear image. With an estimated blur kernel from PanCVPR2016, non-blind image deblurring methods DongECCV2018ZhangCVPR2017PanCVPR2016 generate better results (c,d,e) than the blind method (b). Yet, our DWDN+ method recovers still visibly clearer results (f).
• Figure 2: Deep Wiener deconvolution network. While previous work mostly relies on a deconvolution in the image space, our neural network first extracts useful feature information from the blurry input image and then conducts an explicit Wiener deconvolution in the (deep) feature space through \ref{['eq:deconvolution', 'eq:wiener_filter']}. A multi-scale cascaded encoder-decoder network progressively restores clear images, with fewer artifacts and finer detail. The whole network is trained in an end-to-end manner.
• Figure 3: (a) Blurry image and blur kernel. (b) and (c) show the results of methods that perform the deconvolution in the standard image space and a deep feature space, respectively. (d) Ground truth. The deblurred result (c) from the proposed approach contains fewer artifacts and much more detail in the yellow and red boxes than those in (b).
• Figure 4: Multi-scale feature refinement in our conference version DongNeurIPS2020.
• Figure 5: Example with simulated blur ($1\%$ noise level) from the dataset of SunICCP2013. The result obtained by SchulerCVPR2013 has severe artifacts in (d). For other methods, small-scale structures and detail are over-smoothed as shown in the red and blue boxes of (c) and (e)--(i). Compared to existing methods, our DWDN approach can effectively preserve finer detail as shown in (j). Our improved multi-scale cascaded feature refinement (DWDN+) can make small-scale structures even clearer and sharper as shown in (k).