Improved BM3D method with modified block-matching and multi-scaled images

BM3D-based denoising has been showing high performance in restoring images damaged by additive white noise and there has been intense research on this method and its variants. In this paper, we make three improvements on BM3D (Block-matching and 3-dimensional filtering). Block matching performs poor and affect denoising performance, especially if noise intensity is high. In the paper, we first proposed a new block similarity metric that accounts for characteristic of noise contained in the observed images in order to guarantee accuracy of block matching even in presence of high intensity noise. Second, block size is a crucial hyperparameter for BM3D. The optimal block size varies with the characteristic of images. However, it is difficult to determine such an optimal block size. We proposed a method to mitigate this difficulty in determining optimal block sizes by combining BM3D and multi-scaled images. Finally, in Aggregation of BM3D, the same weight is assigned to every block in three-dimensional structures. In fact, however, the degree with which noise is removed in each block is different. We presented a method of assigning different weights to blocks according to their respective denoising degrees. Experimental results show that the proposed method is competitive with BM3D and even many of state-of-the-art methods. Actually, it brings about 0.1 similar to 0.6 dB pickup in the PSNR (Peak Signal to Noise Ratio) value. Also, we recommend that it may get better results by applying ideas proposed in this paper individually to state-of-the-art methods.