Parallel image de-fencing: Technique, analysis and performance evaluation

Abstract

Image de-fencing is a real-life problem in digital photography, where fences from images are removed in a seamless way such that it appears as if fence was never existed on the image. Often such fence objects are unwanted but unavoidable in real photography. Several attempts has been made to automatically de-fence an image, but the problem with most of the techniques is that the whole image needs to scan many times before Inpainting the unwanted block of pixels, and even it can cause more inefficiency and degrade in performance if the fence detection procedure is automated. Thus there should be a balance between efficiency and accuracy. One way to achieve significant performance gain is through parallelism, but it depends on how well the algorithm lends itself to parallelization. In this paper, we present different parallel versions of a feature based image de-fencing technique that has a lot of potential for parallelism. We implemented a parallel version that uses modern GPUs and NVIDIA CUDA framework to increase the overall efficiency of the algorithm with approximately no noticeable effect on visual quality of the results. We have also determined average speed up of parallel algorithm over serial algorithm through relative experiments.