Temporal noise reduction method for noisy image and related apparatus

Patent metadata

Field	Value
Publication number	US-9721330-B2
Application number	US-201615143190-A
Country	US
Kind code	B2
Filing date	Apr 29, 2016
Priority date	Nov 1, 2013
Publication date	Aug 1, 2017
Grant date	Aug 1, 2017

Abstract

Determining of still/movement may be performed with reference to quantization noise of a first section to which a first pixel belongs, and for different results of determining of whether the first pixel is in a movement area or a still area, different frame difference thresholds applicable to the movement area and the still area are separately set, and different frame difference calculation manners are used, different blending coefficients applicable to the movement area and the still area are selected according to the different frame difference thresholds applicable to the movement area and the still area and the frame difference calculation manners, and a noise reduction blending manner is selected according to the different blending coefficients applicable to the movement area and the still area, the frame difference calculation manners, and a pixel value of the first pixel in a current frame.

First claim

What is claimed is:

1. A temporal noise reduction method for a noisy image, comprising:
acquiring quantization noise of a first section of the noisy image, wherein the first section is any one of multiple sections divided from the noisy image;
detecting, according to the quantization noise of the first section, pixel positions of all pixels in a movement estimation neighborhood in a next frame after the pixels move from pixel positions in a current frame to the next frame, wherein the movement estimation neighborhood comprises neighbor pixels centered around a first pixel, and wherein the first pixel is any pixel in the first section;
determining, according to a pixel position change situation of all the pixels in the movement estimation neighborhood that change from the current frame to the next frame, whether the first pixel is in a movement area or a still area;
selecting a first blending coefficient according to a first frame difference of the first pixel and a preset first frame difference threshold when the first pixel is in the movement area;
calculating a first noise reduction pixel value corresponding to the first pixel according to the first blending coefficient, a pixel value of the first pixel in the current frame, and a movement compensation pixel value of the first pixel in a previous frame, wherein the first frame difference is the difference between the pixel value of the first pixel in the current frame and the movement compensation pixel value of the first pixel in the previous frame, and wherein the movement compensation pixel value is a pixel value of a corresponding position of the first pixel in the current frame and is obtained after movement estimation and movement compensation are performed based on a noise reduction pixel of the first pixel in the previous frame;
selecting a second blending coefficient according to a second frame difference of the first pixel and a preset second frame difference threshold when the first pixel is in the still area; and
calculating a second noise reduction pixel value corresponding to the first pixel according to the second blending coefficient, the pixel value of the first pixel in the current frame, and the noise reduction pixel value of the first pixel in the previous frame, wherein the second frame difference threshold is greater than the first frame difference threshold, wherein the second frame difference is the difference between the pixel value of the first pixel in the current frame and the noise reduction pixel value of the first pixel in the previous frame, and wherein the noise reduction pixel value is a pixel value of a corresponding position of the noise reduction pixel of the first pixel in the previous frame.

2. The method according to claim 1 , wherein acquiring the quantization noise of the first section of the noisy image comprises:
dividing the first section into multiple blocks, wherein each block comprises multiple pixels;
acquiring quantization noise of all the pixels in a first block;
acquiring quantization noise of the first block according to the quantization noise of all the pixels in the first block, wherein the first block is any block in the first section;
separately acquiring quantization noise of all the blocks except the first block in the first section; and
either:
calculating average quantization noise based on the quantization noise of all the blocks in the first section, and using the average quantization noise as the quantization noise of the first section; or
accumulating the quantization noise of all the blocks one by one in the first section, and using the quantization noise that is greater than a preset noise threshold in a cumulative histogram as the quantization noise of the first section.

3. The method according to claim 2 , wherein acquiring the quantization noise of the first block according to the quantization noise of all the pixels in the first block comprises:
determining whether each pixel in the first block is in a flat area;
acquiring the quantization noise of all the pixels in the first block that are in the flat area; and
calculating the quantization noise of the first block according to the quantization noise of all the pixels in the first block that are in the flat area.

4. The method according to claim 3 , wherein determining whether each pixel in the first block is in the flat area comprises:
acquiring pixel values of all pixels in a noise estimation neighborhood, wherein the noise estimation neighborhood comprises neighbor pixels that are centered around the first pixel and that are used to determine quantization noise of the first pixel, and wherein the first pixel is any pixel in the first block;
calculating an edge estimation value of the first pixel according to the pixel values and Sobel edge convolution kernels of all the pixels in the noise estimation neighborhood;
either:
determining whether the edge estimation value of the first pixel is greater than an edge area threshold, determining that the first pixel is in an edge area when the edge estimation value of the first pixel is greater than the edge area threshold; or
determining that the first pixel is not in the edge area when the edge estimation value of the first pixel is less than or equal to the edge area threshold;

calculating a texture estimation value of the first pixel according to the pixel values of all the pixels in the noise estimation neighborhood;
either:
determining whether the texture estimation value of the first pixel is greater than a texture area threshold, and determining that the first pixel is in a texture area when the texture estimation value of the first pixel is greater than the texture area threshold; or
determining that the first pixel is not in the texture area when the texture estimation value of the first pixel is less than or equal to the texture area threshold; and

determining the first pixel is in the flat area when the first pixel meets both of the following conditions: the first pixel is not in the edge area, and the first pixel is not in the texture area.

5. The method according to claim 4 , wherein calculating the edge estimation value of the first pixel according to the pixel values and the Sobel edge convolution kernels of all the pixels in the noise estimation neighborhood comprises calculating the edge estimation value of the first pixel in the manner of EM(x,y)=|NH(x,y)*E_h|+|NH(x,y)*E_v|+|NH(x,y)*E_p45|+|NH(x,y)*E_n45|, wherein (x,y) is a pixel position of the first pixel in the current frame, wherein EM(x,y) is the edge estimation value of the first pixel, wherein NH(x,y) is the noise estimation neighborhood, wherein E_h, E_v, E_p45, and E_n45 are the Sobel edge convolution kernels, wherein * is a convolution symbol, wherein calculating the texture estimation value of the first pixel according to the pixel values of all the pixels in the noise estimation neighborhood comprises calculating the texture estimation value of the first pixel in the manner of Noise_Max_Min(x,y)=Max(abs(value i −value_median))−Min(abs(value i −value_median)), wherein value i εNH(x,y), wherein (x,y) is the pixel position of the first pixel in the current frame, wherein Noise_Max_Min(x,y) is the texture estimation value of the first pixel, wherein value, is an i th pixel value in the noise estimation neighborhood, wherein NH(x,y) is the noise estimation neighborhood, wherein value_median is a middle value or an average value of the pixel values of all the pixels in NH(x,y), wherein determining whether the first pixel meets the conditions: the first pixel is not in the edge area and the first pixel is not in the texture area, comprises determining in the manner of (EM(x,y)≦EGth)&&(Noise_Max_Min(x,y)≦MNth)==1, wherein EM(x,y) is the edge estimation value of the first pixel, wherein EGth is the edge

Assignees

• Huawei Tech Co Ltd

Inventors

• Chen Hai
• Zhang Fengwei

Classifications

• G06T2207/20182

  Noise reduction or smoothing in the temporal domain; Spatio-temporal filtering · CPC title

• H04N19/117

  Filters, e.g. for pre-processing or post-processing (sub-band filter banks H04N19/635) · CPC title

• H04N5/145

  Movement estimation (for video coding H04N19/51) · CPC title

• G06T2207/20021

  Dividing image into blocks, subimages or windows · CPC title

• G06T5/002 Primary

  Physics · mapped topic