A very fast and efficient multistage selective convolution filter for removal of salt and pepper noise

In this paper we propose a multistage selective convolution filter (MSCF) for fast and efficient removal of salt-and-pepper noise (SPN) in digital images. By avoiding the use of order statistics or other computationally expensive procedures, the proposed denoising algorithm is efficiently implemented using convolution blocks, thereby a significant reduction in computation time is achieved. Moreover, in each stage of the proposed structure, a weighted mean filter of an appropriate kernel size is employed to selectively restore a set of noisy pixels qualified by a reliability criterion to improve the performance. The simulation results show that the proposed method denoises much faster than all its competent counterparts, while it achieves a significant performance in both quantitative criteria and visual effects. While noise removal by traditional methods such as AMF takes about 1.092 s and by fast state-of-the-art methods such as NAHAT takes about 0.065 s on each image of the BSDS500 dataset on average, the proposed method dramatically reduces the execution time to 0.005 s.