The Characteristics of Region Gradient for Image Inpainting - Multimedia and Signal Processing - page 129

The Characteristics of Region Gradient for Image
Inpainting

Xueyi Ye, Wentao He, Huahua Chen, and Wangbing Li

The Lab of Pattern Recognition and Information Security,
College of Communication Engineering, Hangzhou Dianzi University,
Hangzhou, 310018, China
{xueyiye,iseealv}@hdu.edu.cn, [e-mail],
[e-mail]

Abstract. The TV model is one of the most classic models in the field of image
inpainting. Having analyzed the computing essences and numerical
implementation process of the TV model, we find that it is easy to resulting in
the edge region fuzzy or fracture of inpainted images using the model.
Accordingly, an improved TV model based on characteristics of region gradient
is proposed in this paper. Here, the region gradient characteristics are mainly
decided by the region direction variable of a pixel and used in the original TV
calculating model. Therefore, the gradient value reflects the image feature more
accurately and the iterative repair makes full use of the neighborhood's
information of a pixel, by which the edge of the missed region is protected and
repaired better. Simulation results demonstrate the validity of the proposed
model by yielding excellent visual effect and showing some significant
performances improvement over the primitive TV model in terms of subjective
or objective image quality measures.

Keywords: Characteristics of region gradient, total variation, digital image
inpainting.

1 Introduction

Image inpainting is the art of filling the missing data in an image or painting with
available information from their surroundings so that the restored image or painting
seems as natural as its original version [1, 2]. Recently, the technology has been
widespread concern of the domestic and foreign experts, some experts have applied
inpainting techniques in de-interlacing [3], image compression [4] and automatic
image recovery [5].
The existing digital image restoration models are separated into two types: the
geometric inpainting and the texture synthesis based on image restoration. At present,
the TV model [6] is one of the most classic algorithms in the field of image
inpainting. However, the drawback of the TV model is poor in large-scale region
inpainting, both the calculation of gradient and the numerical implementation only
used the (, )

xy direction information of image, and ignored the role of several points
in other two directions, which leads the restoration effect is not nature[7].