An Adaptive Directional Non-local Means Algorithm with Size-adaptive Search Window for Image Denoising

The non-local means (NLM) algorithm computes the pixel similarity using a globally-fixed average patch and research window. Few studies have attempted to find a method of adaptively choosing a patch shape and research window size based on local characteristics. To exploit the local characteristics effectively, we propose an adaptive directional non-local means algorithm with size-adaptive search window (ADNLM-AW). The proposed algorithm first computes the direction of each pixel and partitions the entire image region into three classes. Subsequently, the directional rectangular patches are rotated to coincide with the direction of the texture and the search windows are adaptively adjusted to match the local characteristics of the regions. The experimental results show that the proposed algorithm outperforms other current NLM algorithms in terms of visual quality and peak signal to noise ratio (PSNR).