Noise reduction in high-ISO images using 3-D collaborative filtering and structure extraction from residual blocks

This paper proposes a noise reduction method in digital camera images at high-International Organization for Standardization (ISO) setting. Noise characteristics of luminance and chrominance channels in high-ISO images are different, so we decompose an image into luminance and chrominance channels and process them differently. With luminance image, we perform three-dimensional (3-D) collaborative filtering and structure extraction from residual blocks. The proposed method groups similar two-dimensional blocks into 3-D arrays for collaborative filtering in the transform domain. Residual blocks obtained by 3-D collaborative filtering contain structure residual as well as noise. A modified joint image-noise filter is used to extract structure residual from residual blocks. Finally, denoised blocks and structure residual blocks are returned to their original locations. To process overlapping blocks obtained from 3-D filtering and structure extraction, aggregation is used. Chrominance images at high-ISO setting are observed to have coarse-grain noise. To reduce color noise, we use a multiresolution framework, where chrominance images are decomposed into low- and high-frequency subbands. 3-D collaborative filtering is applied to low-frequency subband whereas hard-thresholding is used in high-frequency subbands. Experimental results with various test images at high-ISO setting show that the proposed method gives better subjective visual quality than other state-of-the-art image denoising methods.