A structural refinement method based on image gradient for improving performance of noise-restoration stage in decision based filters

In recent years, decision based filters (DBFs) are the most popular technique for impulse-noise restoration. The DBFs consist of two stages: noise-detection and noise-restoration. The performance of noise-restoration stage affects the quality of DBFs significantly. In this paper, we presented an effective structural based refinement method which could be adopted as a complementary stage after DBFs to improve the quality of the final restored image. Here, we assume that the preliminary DBF has detected the noisy-pixels and has restored the intensities of the noisy-pixels. In our proposed refinement method for each detected noisy-pixel, based on local structural information of the image, the previously restored intensity of noisy-pixel is modified more accurately. This is performed by analyzing the gradient of output restored image of preliminary DBF and calculating direction of contour which are passed through the noisy-pixels. Then based on the angular difference of contour-direction with 4 main lines, which are passing through the noisy-pixel, the previously restored intensity of noisy-pixel is replaced with weighted means of surrounding pixels' intensities. Since the structures in images are more recognizable for low-density impulse-noise, our method is more effective in this case, however a small improvement is obtained for high-density impulse-noise.