237 2014 Fourth International Conference on Emerging Applications of Information Technology An Adaptive Bilateral Filter For Inpainting A Dao Nam Anh Department of Information Technology Electric Power University 235 Hoang Quoc Viet road, Hanoi, Vietnam Fig. 1. Example of inpainting: a) Input image, b) Mask and full/partial Gaussian filter, c) Inpainting result Abstract— An adaptive model of bilateral filter is presented for digital inpainting. The model works by transforming inpainting into an equivalent energy condition minimization and generation of patches for missing areas by interpolating within working frame. It combines knowledge of local structure by bilateral filter and intensive value. Bilateral filter is adapted to missing regions to check similarity of regions to fill-in. Standard deviation in range kernel of the filter is regulated by total variation. This helps to create patches that keep edges. Total variation is also efficient for detection of missing pixels which possibly stay on edges. Benefit of the model was demonstrated in experiment of inpainting for gray and color images. Keywords— Inpainting, total variation, bilateral filter, similarity, edge-preserving filter I. INTRODUCTION Image inpainting play a fundamental role in improvement for unity and completeness of image. It is still open research topic in image analysis. Typical algorithm would take missing areas and use information of good areas to restore the missing areas. The conventional methods for solving this problem use Partial Differential Equations (PDE) in iterative algorithms. Pixels of missing area in the border with good areas are checked with their neighbor pixels in good area by PDE based functions. This runs in each iterative cycle until all missing pixels were recovered. Exemplar based methods, similar to PDE approach, also run around the border of missing area, but they check each small region of pixels but not a single pixel. Exemplars with specific structure are created by small regions in known area and they are used like example to fill-in missing area. Bilateral filter introduced firstly into image denoising application. This non-linear filter uses Gaussian in the spatial domain where weight of each pixel is influenced by the intensity domain. Its advantage is to eliminate noise but keep edges. This paper work follows concept of the exemplar based methods, incorporating with total variation (TV) and bilateral filter to check similarity of structure. In adopting bilateral filter concept, incomplete bilateral filter was introduced in optimal condition, presenting similarity of intensive value and structure between missing and good areas of image. II. OUTLINE OF PAPER Related work contributions and notation for this work are presented in section 2, 3, 4, 5. In section 6, major concept of adaptive bilateral filter for inpainting is formulated. The section explains how bilateral filter can be transformed into suitable form for checking similarity regions in image. Section 6 presents the experimental evaluation of the total variation bilateral filter for inpainting gray and color images. Before describing paper work in details, let’s see figure 1 for visual display of working concept. Fig.1a shows input image. Fig.1b draws full Gaussian net in left and partial net in right. The partial net is an example of incomplete bilateral filter for a pixel on border of paint mask, that produces result in fig.1c. III. PREVIOUS AND RELATED WORK Image inpainting has practical application and several researches addressed to the topic. PDE based algorithm is proposed by Marcelo Bertalmio et al [1]. PDE based methods found success in iterative algorithms. Pixels of missing area in the border with good area are checked with their neighbor pixels in good area by PDE based functions. This is run in each iterative cycle until all missing pixels were recovered. Exemplar based methods, similar to PDE, also run around the border of missing area. It checks each small region of pixels but not a single pixel. Exemplars with specific structure 978-1-4799-4272-5/14 $31.00 © 2014 IEEE DOI 10.1109/EAT.2014.13