Inverse dynamical non-local filter

Estimation of saliency requires multiple techniques and often results in unacceptably poor match quality of saliency map construction. This work addresses the challenge by introducing the inverse dynamical non-local filter and automated analytical saliency estimation method by the filter. The inverse non-local mean filter is a new implementation of the non-local means filter which is a practical denoising method. In contrast to the traditional application, the inverse filter represents the dissimilarity of neighbors of the target and the source pixels. An adaptive version of the filter with dynamical radio of search window and the degree of filtering is presented for predicting image regions which grab the most visual attention. Quality metrics are examined for a saliency benchmark and compared to other methods. Experimental results demonstrate the effectiveness of the method and state-of-the-art performance.