BM3D denoising based on minimum GCV score

Block-matching and 3D filtering(BM3D) is an effective denoising approach introduced by Dabov et al. which utilizes 3D transform collaborative filtering on small similar patches extracted from the image corrupted by additive white Gaussian noise. However, when it comes to blind filtering, the denoising performance will worsen rapidly. In this paper, an improved version of BM3D which applies the adaptive noise estimation method based on minimum generalized cross-validation (GCV) score is proposed. Firstly, the noise standard deviation is estimated by minimizing GCV score. According to the optimal smoothing parameter setting, curve fitting is then analyzed to build a formula to modify the estimated noise level. The modified smoothing parameter computed by the built formula is used in BM3D filtering, which induces BM3D algorithm to become adaptive to variant noise. Experiment results display that the proposed method outperforms the original BM3D algorithm in terms of the visual effect and the image quality.