Raw data denoising (ii)--BM3D algorithm of ISP module

Before formally starting this article, let's review some basic ideas and methods of CFA image denoising. Then I will be in detail with you to share their learning understanding of the BM3D algorithm, operation process, its advantages and disadvantages, the final will give the algorithm for reference.

In the ISP module, the researchers will discuss whether the Denoising module (Noise Reduction) is either before or after the Mosaic module (DEMOSAIC). If before, with the denoising process, the noise point is eliminated, along with the loss of color information, if after that, the complex interpolation process will change the statistical model of noise, making it complex and difficult to calculate. Therefore, it is more the case to choose to do the demosaic before the noise.

The CFA (Color Filter Array) data cannot use the traditional grayscale image denoising algorithm, because the adjacent pixels in the CFA image have different color information measures, the block structure of the CFA image is not in the traditional sense of smoothness and the piecewise constancy, So that the general denoising algorithm does not apply to the CFA image. CFA data is also not able to use the color image denoising algorithm because each pixel contains only one color channel information.

① One method is to divide the original CFA image array into four small pieces (r,g1,g2,b), respectively, using the gray-scale image denoising method for the four blocks. This approach often behaves poorly because important color correlation information is overlooked. The CFA denoising algorithm can improve its effect by utilizing the spatial and color correlation of CFA data.

② Another method is to construct a low-resolution RGB image using the information of each color block in the CFA image, which makes good use of its color-related information, but does not protect the high frequency information in the space domain.

③BM3D (Block Matching-filtering algorithm) algorithm is proposed, by restricting the image block has the same color configuration structure to achieve the purpose of processing the CFA image.

The following is a detailed description of the BM3D algorithm:

1. Basic estimates

1). Block-wise Estimation (Block-wise estimates)

Grouping (Grouping), find all the blocks that are similar to the currently processed image blocks, and pile them together to form a 3-D array (grouping).

Union hard threshold (collaborative hard-thresholding). The 3D transform of the grouped group is done, the hard threshold 3D transform coefficient achieves the purpose of weakening the noise, and then the 3D inverse transform is taken back to get the post-denoising estimation of the image block in the packet. and return to where they were before.

2). Aggregation (Aggregation) The weighted average of the pixel points of the estimated image block, resulting in the final pixel value, which is the final base estimate.

2. Final estimate

1). Block-wise Estimation (Block-wise estimates)

Grouping (Grouping), using the method of image block matching, we find the original noise image and all the blocks similar to the current processing image block in the basic estimation image, forming two 3-D arrays (groupings).

Combined Wiener Filter (collaborative wiener-filtering). The 3D transform of the two groups that have been organized, the energy spectrum of the basic estimation image as the real energy spectrum to the noise image packet Wiener filtering, It then goes back to the 3D inverse transform to get all the grouped image block estimates and return to where they were before.

2). Aggregation (Aggregation)

A weighted average of all the pixel points that have been repeatedly obscured by the estimated image block is obtained, resulting in the final pixel value, which is the final result of the final estimate.

Bm3d how to find similar image blocks organization 3D grouping: As shown, the squares are shown as individual image blocks, and the image blocks selected in the upper-left groupings have corner (cusp) features, and other groupings are similar.

How Bm3d is handled in the CFA: as shown, the image blocks on the left side of the packet have different color configurations, that is, the r,g,b arrangement is not consistent, and the right side has the same color configuration, so it is very reasonable for processing the CFA image.

The above will bm3d the principle of the algorithm and details of the operation are introduced to you clearly, the following to see the bm3d for the CFA image algorithm effect:

Original noise image

Median filtered image

Bm3d Noise removal image

The algorithm effect (PSNR) is obviously better than median filter, the denoising module ends, and the next one starts to introduce two kinds of color enhancement algorithms.

Bm3d Online has ready-made MATLAB code, algorithm link URL: http://www.cs.tut.fi/~foi/GCF-BM3D/

Reference documents:

1). A case for denoising before demosaicking color fiter array Data,sung Hee Park etc.

2). A Framework for wavelet-based analysis and processing of the color filter array images with applications to denoising and demo Saicing,hirakawa etc. Havard University.

3). Cross-color bm3d Filtering of Noise Raw Data,aram Danielyan etc.

Raw data denoising (ii)--BM3D algorithm of ISP module