Exemplar-based Colour Image Inpainting: A Fractional Gradient Approach with Improved Confidence Term

Abstract

The reconstruction of visually persuasive image from the scratched or teared images is known as image inpainting.It can also be used for intentionally removing some data from the image which is not required but got captured by accident. This prediction of data in place of removed data is done by various inpainting techniques. But the images recovered must look authentic to the user. Texture synthesis and inpainting is combined in the exemplar based methods to fill the void in image, so that it looks pleasant to viewer’s eye.

In exemplar-based technique, two types of terms are used namely confidence term and data term. A new priority function in additive form is used for the calculation of prioritisation of pixels to be inpainted in the target region, with an improved confidence term and a fractional based gradient function is used for defining the data term which is more efficient, instead of traditional multiplicative form because the latter causes the priority value to decline rapidly and resulting in inefficient filling of patches.

The proposed algorithm is tested by removing text written on real world images and removing a certain object from the image so that resultant image looks authentic and results are found to be better and more efficient than the present traditional techniques in terms of performance metrics MSE, PSNR and SSIM. From Criminisi Algorithm, there is an increase of PSNR ranging from 0.24 dB to 7.90 dB and from Chandersekran method, increase of 0.1 dB to 0.96 dB is achieved. Also the final image looks more legitimate to the human eye as it should not appear to be tempered with.