Image inpainting is considered as a predictive process to compute the missing image data without ... more Image inpainting is considered as a predictive process to compute the missing image data without introducing any undesirable artifacts. Most of the existing methods in the literature for this purpose work very well for small regions but introduce blur for large holes. In this paper the work done is an extension of a previous work where the novel framework for affine and flip invariant inpainting of color images was proposed which combines structural similarity index measure, an improved version of color angular radial transform, frequency domain based image registration and Kekre's LUV space based blending. We now extend it by including histogram of oriented gradients to preserve the texture and parallelizing the implementation using OpenCV on multicores and GPUs. The parallel implementation of the original algorithm achieved a speed up of 11 on multicores and up to 13 on a GPU and around 18 on the multicores and a GPU hybrid platform. The experimental results are found out to be more computationally efficient, fast, and more visibly plausible.
Image inpainting is considered as a predictive process to compute the missing image data without ... more Image inpainting is considered as a predictive process to compute the missing image data without introducing any undesirable artifacts. Most of the existing methods in the literature for this purpose work very well for small regions but introduce blur for large holes. In this paper the work done is an extension of a previous work where the novel framework for affine and flip invariant inpainting of color images was proposed which combines structural similarity index measure, an improved version of color angular radial transform, frequency domain based image registration and Kekre's LUV space based blending. We now extend it by including histogram of oriented gradients to preserve the texture and parallelizing the implementation using OpenCV on multicores and GPUs. The parallel implementation of the original algorithm achieved a speed up of 11 on multicores and up to 13 on a GPU and around 18 on the multicores and a GPU hybrid platform. The experimental results are found out to be more computationally efficient, fast, and more visibly plausible.
Uploads
Papers by karthik p