CONTEXT AWARE EXEMPLAR-BASED IMAGE INPAINTING USING ADAPTIVE IMAGE DIVISION

Abstract

Image inpainting is the process to reconstruct the missing or corrupted regions of the images. There are many applications of image inpainting algorithms such as object removal from images, scratches removal form old photographs for restoration, text removal from image, missing blocks recovery in image encoding and decoding etc. Various algorithms have proposed for image inpainting but there is no 'perfect' inpainting algorithm, each technique has some advantages and limitations. In this thesis exemplar based image inpainting techniques have studied extensively, their limitations are found and improvements have proposed. In existing inpainting techniques, complete image is searched for filling the missing region. Contextual information could be used to limit the search space to similar regions. Therefore context aware exemplar based image inpainting using adaptive image division is proposed. in the proposed method, image is divided into contextually similar regions of adaptive sizes. Instead of complete image, proposed algorithm uses combination of contextual similar blocks as a search space for each best-match search. A new priority term is also introduced to define the order of searching for best-first patch. Experimental results on benchmark images demonstrate the improvement and significant acceleration of existing exemplar based method. In the proposed method we need to perform extra step of image division than existing exemplar based technique. But in proposed algorithm, searching takes very less time than existing method to give reduced total computation time. Quantitative evaluation of the proposed algorithm is performed by using the Peak Signal to Noise Ratio (PSNR) as a measure. Various experiments have performed by imputing the noise in test images. Original images are used as ground truth and corrupted images are used as input to the inpainting algorithms. We have compared the results of three inpainting techniques: Criminisi method, Content aware fill of Photoshop and proposed method by computing the PSNR values between the obtained results and the ground truth images. We have observed that proposed method gives better results with higher PSNR values than other methods. The effect of image size is observed by performing experiments with varying size images. Experimental results have shown that increment in computation time in proposed method is much less 75.9% less than existing method (2389.1 sec. in proposed method and 9519.9 sec in existing method) for increment in image size of 47.19%.