IMAGE DECOMPOSITION USING DECONVOLUTION Sunghyun Cho, Hyunjun Lee, Seungyong Lee POSTECH ABSTRACT We present a novel method for decomposing an image into base and texture layers. Our method is simple and effective, and can handle textures of high contrast, which traditional image filtering techniques may not handle efficiently. The method first removes high-frequency texture information us- ing low-pass filtering, and then restores structural information of the image using a deconvolution operation. Experimental results demonstrate the effectiveness of our method. Index Terms— texture layer, image decomposition, im- age filtering 1. INTRODUCTION An image can be decomposed into base and texture layers. A base layer consists of smoothly varying regions, and con- veys large-scale structural information of an image. A tex- ture layer depicts small-scale variations, and contains details of image appearance. Since these two layers provide differ- ent types of information, image decomposition helps solving many kinds of problems in computer vision and graphics, e.g., segmentation, object matching, and non-photorealistic image abstraction. In spite of its benefits, decomposing an image into base and texture layers is still a challenging problem. It is under- constrained as the number of unknowns is twice the number of pixels. Distinguishing image structures from details is in- herently an ambiguous task. In this paper, we present a simple and effective method for decomposing an image into base and texture layers. The key insight of our method is that low-pass filters can effectively remove textures. Texture can be defined as small-scale oscil- lating patterns and local variations of image intensities. Un- der this definition, textures intrinsically correspond to high- frequency components of an image, which can be easily re- moved using a low-pass filter, e.g., a Gaussian filter. However, applying a low-pass filter to an image also blurs structural information, such as edges and region bound- aries. To recover these structural information, we perform This work was supported by the IT R&D program of MKE/MCST/KEIT (2008-F-031-01, Development of Computational Photography Technologies for Image and Video Contents) and the ERC Program of MEST/NRF (R11- 2008-007-01002-3). deconvolution on the low-pass filtered image. The deconvo- lution operation can restore the boundaries between regions of different textures because such information still remains in the blurred image although it has been tampered. On the other hand, high-frequency texture components have been completely lost by the low-pass filtering and are not restored during the deconvolution. The deconvolution result provides the base layer, while the texture layer can be obtained by subtracting the base layer from the given image. Our method can handle textures of high contrast effec- tively, while preserving boundaries between regions of differ- ent textures. It is simple and fast, consisting of simple low pass filtering and deconvolution steps. Both steps run fast and are easy to implement with plenty of resources including source codes available on internet. 1.1. Related Work Bilateral filters [1] and feature-preserving smoothing opera- tors have been used to decompose an image into base and de- tail layers [2, 3]. However, textures often have high-contrast edges and bilateral filters are not effective to handle such tex- tures. Farbman et al. [4] introduced a weighted least squares (WLS) based filter, which better preserves multi-scale fea- tures than bilateral filtering. However, their method cannot smooth out textures of high contrast, similarly to bilateral fil- ters. Besides filtering based techniques, total variation (TV) minimization based techniques have been developed for im- age decomposition. Rudin et al. [5] introduced a TV mini- mization based texture removal method. Vese and Osher [6] extended it using the space of oscillating functions introduced by Meyer [7]. Yin et al. [8] used an L1-norm fidelity term with total variation regularization to decompose an image into base and texture layers. However, these methods produce quantization artifacts or do not completely remove textures from a base layer (Sec. 3). In the sense of “blur-and-deconvolve”, filtering by recon- struction technique is similar to ours. It first creates a marker image by applying a low-path filter then reconstruct the im- age. Maragos and Evangelopoulos [9] proposed a multiscale leveling based technique. Wilkinson showed interesting re- sults by applying levelings with reconstruction criteria [10].