TW200826644A - Method of image noise reduction - Google Patents

Abstract

A method for reducing image noise includes calculating the mean pixel-correlation between the central pixel and its neighbor pixels within a working window in an image. Besides, the method includes calculating the weighted horizontal similarity between the central pixel and horizontal neighbor pixels and the weighted vertical similarity between the central pixel and vertical neighbor pixels within a working window. Then, the method is to judge if the central pixel is a noise, followed by noise reduction process if it is.

Description

200826644 IX. Description of the Invention: [Technical Field] The present invention provides a method for attenuating image noise, in particular, a method for calculating pixel correlation between a pixel at a center point and an adjacent pixel, and then selecting a size of the working window to attenuate image noise And a method of attenuating image noise by using a weighted similarity between a center pixel and an adjacent horizontal and vertical pixel. [Prior Art] In the era of multimedia communication, video information plays an important role. However, even if the camera specifications are perfect, no image is absolutely perfect, and it is expected to be disturbed by the presence of noise. The main source of noise in digital images is the image capture (digital) transmission. The performance of the imaging sensor is affected by many factors, such as the environment during the film (four), and The sensing element itself, for example, on the 掏(10) camera capture image, the luminosity and the sensor's degree are important factors affecting the amount of noise in the generated image. The main reason for the image being transmitted is the transmission channel. Interference, such as a charged particle interference using a wireless network pass. Like the possibility of being filtered by lightning or other atmospheric digital images are in the protection of the shadows of the actor Tianshan like fine instant mitigation of noise, is a part of the processing like the necessary & 2 ^ 〗 in many applications based on computed images In the application of the game, the m. Any noise will cause serious entertainment results. Therefore, the noise attenuation is more important, and the visual quality is improved. Cutting, identifying or interpreting performance tasks. For example, encoding, analyzing, and cutting 200826644 μ 2 pixel material will be interfered by noise due to (4) _ _ == poor shadow money taking conditions or encountering errors in image transmission; pulse noise is very easy to be It is perceived by the eye and causes some serious errors in image processing applications. Therefore, pulse noise sag is usually used as a pre-processing in some image processing systems, such as image quantization,... Wait. An optimal pulse noise filter must smooth out dissimilar pixels in the homogenous region, preserve edge information and not change any natural image information. Different pulse noise filtering algorithms have been published in the past few years and their The purpose is to protect the image details while filtering the pulse noise. Some typical nonlinear filters, such as the median filter and the intermediate weighted median filter, can attenuate most of the pulse noise and will Most of the image details are filtered out. For the application of reducing image noise, many scholars have proposed some methods. The following are related references: [1] Tao Chen and Hong Ren Wu, " Space Variant Median Filters for the Restoration of IinpulsG Noise Corrupted Images,', IEEE Transactions on Circuits and Systems 11:

Analog and Digital Signal Processing, Vol. 48, Issue 8, Pages: 784 - 789, Aug. 2001.

[2] I. Aizenberg and C. Butakoff, “Efficient Impulse Detector Based on Rank-Order Criteria”, IEEE Signal Processing Lett., Vol. 11, Pages: 363 - 366, Mar. 2004. 6 200826644 [3] G· Pok, Jyh-Charn Liu, and A·S·, “Selective Removal of Impulse Noise Based on Homogeneity Level Information,,, IEEE Trans. Image Processing, Vol. 12, Pages: 85 - 92, Jan· 2003 [4] XD Jiang, “Image detail-preserving filter for impulsive noise attenuation '', IEE Proceeding-Vis· Image Signal Process, Vol. 150, June 2003.

[5] X. Xu, E. L. Miller, Dongbin Chen, and M. r -

Sarhadi, "Adaptive two-pass rank order filter to remove impulse noise in highly corrupted images '', IEEE Trans Image Processing, Vol. 13, Pages: 238 - 247, Feb. 2004. In reference [2], I Aizenberg and C. Butakoff proposed the Differential Rank Impulse Detector (DRID) to effectively detect pulse noise. In a working window, the order of the pulse noise and the difference between the center point image and the prime order are very different. Large, the median of different sequences is always in the middle' and the pulse noise is near the ends, thus obtaining a simple pulse noise detector with the idea of placing the position of the pixel of interest with the criticality. The value is compared and can be expressed as the following expression: (Electrical J y) V +1 ; where Xi, j is the center point pixel of a working window, and R ( xi, j ) is

Xi, j is the order after sorting, N is the number of pixels in the working window, and s is a critical value. 7 200826644 This method is a method that can easily determine whether it is interfered by pulse noise. It can be a good result, but the situation of misjudgment is quite large and cannot be guaranteed. Whether a pixel is interfered by pulse noise, and if Pixels are not disturbed by noise and their order is near the ends of the noise. In order to overcome this question, not only should the test order be considered, but the value can be expressed as another expression: &gt;(# —calling))Λ(( ^) · Its t, dl, rj can be expressed as the following formula: , ifR{XiJ)&gt;MEDij 0 •-~team J+l],ifi?(' y·) <娜/y, else

Var ( k )疋 is a grayscale value sorted by k. This detector is based on comparing the position and absolute value between pixels in a working window, providing an effective speed, 'has smooth image and can be applied to other Any filter method. In reference [3], G·Pok, Jyh-Charn Liu, and A. s. Nair propose a conditional signal adaptive median filter (c〇nditi〇nal

Signal Adaptive Median Filter (CSAM), based on judgment, median value; the filter is mainly composed of two functions, a necessary condition for judging and filtering noise, wherein the first function is used to determine Whether there is noise in a working window, the second function is used to smooth the pixel values of the noise. ‘ The algorithm of this method is as follows. Step 1. Calculate the upper and lower limits of the homogeneous region. Step 2: Detection of pulse noise: 8 200826644 In a 3x3 work full of gentlemen x|8, the center pixel is xO, 8 neighbors are small=1, and ch is 8 neighbors. • The number c 1 of the H-pixels that is homogenous to the center-point pixel xO is the number that is not the same as the center-point pixel χ〇. Right ch> ci, the center point pixel is χ〇 as a signal; if, then the center point pixel is χ〇 as a noise candidate. Step 3: Refining the selected pulse noise: In order to minimize the error of the debt measurement, they use different wave methods to remove/from % of noise interference pixels from the noise candidate set. Most of the detected pixels are located near the edges and in the image details. The two-pixel knives are two groups: one group is similar to the center point pixel, and the group is not similar to the center point pixel, if the number similar to the center point pixel is larger than the center point pixel (four), then This center point pixel is treated as a signal 'and moved from the noise candidate set ^, this step - straight execution until the number of noise candidate sets is no longer reduced. Step 4: Use the median filter to filter out the noise. · If the number of pixels similar to the center point is less than 3 in a 3x3 working window, use the 3x3 median filter to filter out the noise; otherwise, Use a 5x5 median filter to filter out noise. The purpose of this method is to achieve near-perfect pulse noise detection, and the results after restoration have excellent visual quality. α In reference [4], χ·D· Jiang proposed a Truncati〇n filter in which the grayscale value of one pixel (i, 』·) is x(i, 〗), and N sizes can be found as MxM. The square window contains this pixel. This window 9 200826644 is called an internal window and is represented by wik. For each internal window, there is a corresponding external view W0K whose size is (M + 2r)x(M+ 2r), r^ M , J is ', (7), defined as the same center as the internal window. Point, in this way, you can find a closed surrounding zone with βκ, κ=ι··· Ν, thickness r, and closed surrounding band ΒΚ defined as βκ= w〇K—ψικ. Let ~ and 々 denote the maximum and minimum grayscale values in each closed band, and then use the maximum or minimum value of ° and its surrounding enclosing group to determine whether there is interference from noise. The purpose of this method is Image detail can also be protected when attenuating noise. In reference [5], X·Xu, Ε·L·MiUer, D·(3)(5), and Μ· Sarhadi propose an adaptive two-stage median filter (^叩衍)

Two-Pass Median Filtering (ATpMF), when a chopper with a high noise ratio (such as a median chopper), may produce an effect of dissatisfying μ. Performing this filter twice can give better results, called two-stage. The method is dedicated to two goals. First, the algorithm using a two-stage sequential arrangement of the waver can arrange the filter benefits in a high noise ratio in a higher order; consider more noise, the first =, use Estimating the space knife of the pulse noise It shape and correcting the error caused by the first filtering operation' The method is as follows: v The median filter is used to filter out the image noise and obtain an estimated spatial distribution. The size of the situation and the pulse noise value. V. First, after judging the noise filtering by step one, which pixels are 匕 and correct, these pixels are replaced by the original pixels, and remain unchanged in step 3 200826644. Step 3. Use the median filter again to filter out image noise. The purpose of this method is to attenuate images that are subject to high noise ratios of pulsed noise and can be applied to any sorted sequence filter. It can be seen that many algorithms for attenuating image noise have been proposed in the prior art, but some algorithms may only be applied to images with high noise ratio and pulse noise interference, and in some cases, may occur. In case of misjudgment, and in addition to effectively attenuating noise, it is necessary to pay attention to the details of the protection image. SUMMARY OF THE INVENTION The present invention is directed to an image noise attenuation method that is designed to provide better noise attenuation. In order to achieve the foregoing object, the present invention provides an image noise attenuation method for calculating an average pixel correlation between a center point pixel and an adjacent pixel in a working window, and summing the average pixel correlation of all pixels. 'Get the total pixel correlation of the entire image. Determining whether the difference between the total pixel correlations of the two adjacent threshold values is less than a first preset value, and if the difference between the total pixel correlations of the two adjacent threshold values is less than a first preset value, recording the total Pixel correlation. Determining whether the total pixel correlation is greater than a second preset value, and if the total average pixel correlation is greater than the second preset value, using a small working window to attenuate image noise. In addition, the method includes using a large working window to attenuate image noise if the average pixel correlation is less than the first preset value, and calculating a vertical direction of a weighted central pixel and a weighted vertical neighbor pixel in a working window. 11 200826644 Weighted similarity and horizontally weighted similarity of weighted horizontal neighbor pixels. Determining whether the two weighted similarity is less than a third preset value, and if the two weighted similarities are less than the third preset value, retaining the center point pixel, if one of the weighted two similarities is greater than the first For three preset values, use the working window to perform a median filter to attenuate image noise. The above image noise attenuation method can mainly achieve the following effects: 1. Improve the effect of image noise attenuation: The invention can effectively select the appropriate size window to improve the image noise attenuation effect. 2 fast and surely attenuating image noise: the present invention determines the average pixel correlation of the center point pixel and its neighboring pixels in the working window of appropriate size, and the center point pixel is adjacent to its adjacent vertical and horizontal directions The weighted similarity of the pixels, which in turn quickly and reliably attenuates image noise. [Embodiment]

The invention is an image noise attenuation method, which can be divided into several levels. Firstly, it is discussed whether the pixel correlation in the image is interfered by the interference of the noise based on an ηχη working window, and the present invention defines the pixel correlation ^\ Xo^k\&lt;T 〇, otherwise ,l^k&lt;n2 -1 and 1 &lt;

i&lt;N where ~ represents the center point pixel in the working window, A represents the working window /, his pixel, that is, the neighbor pixel of the center point pixel %, N represents the number of pixels, and T is customized by the user The preset value of the working window 12 200826644 can also be the working window of other geometric shapes. The result obtained by the above equation is 丨, which represents that there is pixel correlation between the central point pixel % and its neighbor pixel. Otherwise, the center point pixel % has no pixel correlation with its neighbor pixel A. The present invention further accumulates the pixel correlations in all the working windows, and finally divides the total number of pixels to obtain the total pixel correlation, as shown in the following equation:

IN — ^^LCik , hk S n2 -1 and hn n After this statistic, Gq will be between 〇 and 丨 (0%~100%); as the threshold is set larger and larger, GQ The percentage will also get higher and higher. »月&gt; As shown in Figure 1, the present invention selects two different types of images: L_(a) and Lizard(b), and calculates the total pixel correlation of a central point pixel and an upper neighboring pixel, a preset value. τ is set to 〇 to 5〇, and the obtained result is as shown in Fig. 2; the present invention adds 20% image pulse noise to the [(4)(4) and 仏(4) image tunneling machines (please refer to Fig. 3) The pixel values of the disk are all distributed between G and 255, and the total pixel correlation of the center point and the \3 pixel is calculated. The preset value T is set to 0 to 5 = the result of the service is shown in Figure 4. Shown.

The %/2 graph less than two and the curve in graph 4 can find that the preset value T is flat, and the increase in the total pixel correlation is the largest, and t is more in the smooth ε field. Once, "more, the first reading of the first reading \ pixel correlation is the most obvious increase (please = above and the third picture), with the increase in the default value of the association is also slower... as the average / w is reduced, extended The phenomenon clearly indicates the ratio of the horizontal/monthly region and the edge and non-smooth regions in the image of 200826644, which is the main factor that causes the increase of the total pixel correlation. ^ Soil in ~ like the noise attenuation needs to save a large number of images The original natural resources - this &amp; month reference total pixel correlation increase, for each image set a preset value Τ (hereinafter referred to as the _ preset value), when the total pixel relevance of the towel field is smaller than the first Set the value Ti to record the total pixel correlation (LCikm, , Figure 5), as shown below: ifferent ^ '(LCik) j^X 1 &lt; k &lt; n2 -111 &lt; i &lt; N And 0 &lt;j&lt;255, and then determine whether the total pixel correlation is greater than another preset value k (hereinafter referred to as ^ ° If the total pixel correlation is greater than the second preset value Tc, use the small window to attenuate the image noise to reduce the image noise. Otherwise, use the large working window to invent the blur method to detect the image pulse noise, including: The farther the distance between a central point pixel and an adjacent pixel is, the lower the total pixel correlation is, the lower the value is, the lower the total pixel correlation is, the lower the Euclidean distance is. * Therefore, the present invention is based on the length of the Euclid distance between a central point pixel and an adjacent neighbor, and the reciprocal of the Euclidean distance is the weighted value of each pixel, as shown in the following equation.

L i in A represents the Euclid distance.

Determining the weighted center point de I", the two similarities between the pixel and the weighted level and the weighted vertical adjacent pixels are not less than a preset value Ta (hereinafter referred to as a third preset value), as follows As shown in the figure, · 14 200826644 D = ^TjWix〇-^j^iXi ^ fl5 [0, otherwise If "is equal to the center point pixel is the original natural information, otherwise the center point pixel is contaminated by image pulse noise, use The median filter attenuates image noise. . For example, as shown in FIG. 6 , the present invention uses a diamond type working window to reduce image noise ' Χ 0 is a center point pixel, and XI to XI 2 are adjacent pixels of a center point pixel Χ 0. The method of pulse noise detection is as follows:

Dh Η3 χ χ〇- 〇·5 x W Χ7 - 0·5 X Χ8| pv = |3 x X〇- 0.5 x Xi - X3 - Xi〇- 0.5 x Χη| If the Dh ' Dv is less than the third pre- When the value Ta is set, the center point pixel is the original natural information, otherwise the center point pixel is dyed by the image pulse noise, and the median filter is used to attenuate the image noise. Please refer to Figure 7, the flow of image noise attenuation method 7〇&amp; contains the following steps: Dream Miscellaneous 70 2 ··Compute the total pixel correlation of two adjacent thresholds (Pc 1, PC2) 0 Less Step 704: Calculating the difference between the total pixel correlations of the two adjacent thresholds (PCD) ° Dream 706 ··Determining whether the difference in total pixel correlation (PCD) is less than or equal to the third preset value (Ti ); less than or Equal to the first preset value (Ti ) to record the total pixel correlation (PC1) ° Less 708: Determine whether the total pixel correlation (PC1) is less than the second pre 15 200826644 set value (Tc ); less than the second preset The value (Tc) uses a large window of window attenuation to affect the noise. Step 710: Determine whether the total pixel correlation (pcl) is greater than or equal to a second preset value (Tc); and greater than or equal to the second preset value (Tc) to attenuate image noise using a small working window. Step 712: Calculate the weighted similarity (Dh) of a center point pixel and a horizontal neighboring pixel and the weighted similarity (Dv) of a center point pixel and a vertical neighboring pixel. Step 714: Determine whether the weighted similarity (Dh) of the horizontal neighboring pixels and the weighted similarity (Dv) of the vertical neighboring pixels are all less than or equal to the third pre-existing value Ta, if the weighted similarity (this) and vertical of the horizontal neighboring pixels The weighted similarity (Dv) of the neighboring pixels is less than or equal to the third preset value Ta, and the center point pixel value is directly output. Step 716: Whether the weighted similarity of the adjacent horizontal pixels (Dh) and the vertical neighboring pixels (Dv) is greater than the third preset/value if the weighted similarity of the horizontal neighboring pixels (Dh) is perpendicular to the adjacent pixels. The weighted similarity (Dv) is greater than the third preset value Ta, and the center point pixel value is output after passing through the median filter. Referring to Figure 8, in the test image of Figure a of Figure i, random generated pulse noise is added, and the noise ratio is 5〇%, respectively. In all tests, random pulse is used. For the noise, the pixel value is evenly divided: between the basin and the 255, and the test image obtained by the present invention is referred to in the ninth filter, and the ATW filter is described in the reference (1). The chopper/chopper is a filter proposed by the aforementioned reference (2), and the CSAM filter is a filter proposed by the aforementioned reference (8). The MSM filter is filtered by the aforementioned reference (1). Device. Please refer to Figure 10, which adds randomly generated pulse noise to the test image. The noise ratio is 5% to 5 (10) respectively. In all tests, random value pulse noise is used, and the pixel values are all The sentence is distributed between 〇 and '' and the test image obtained by the present invention is referred to in Figure ii. Lizhong, A Cong filter, wave filter is the filter proposed by the aforementioned reference [5], two... n. The filter is the filter proposed by the aforementioned reference [4]. The DR ID filter is a filter proposed by the aforementioned reference of the Yanjiang River [2], and the CSAM data is referenced by the aforementioned reference. The stitcher proposed in [3], _filter ^ is a filter proposed by the aforementioned reference [1]. In addition, the present invention defines a novel pixel correlation, and estimates an appropriate size of the working window to attenuate image noise, and the second window can be applied to other image noise attenuation/amplitude images. The efficiency of noise reduction. The image noise attenuation method according to the present invention can be used in an image processing process, and the present invention can be stored in a computer readable medium, a first disk, and the like, Saman. The media "woman skirt" to the computer system can realize the image noise attenuation method of the present invention, which can also be realized in a product road, and can be integrated into various devices through the integrated circuit. The method of image noise attenuation of the present invention can be carried out. - 17 200826644 The above-described embodiments are only for explaining the present invention, and are not limited to the scope of the present invention, and the parameters mentioned in the text, such as the first The preset value Ti, the second preset value Tc, and the third preset value Ta are not limited to a fixed value, and an optimal value may be selected according to different image characteristics to achieve better noise attenuation performance. The working window mentioned is a diamond type working window, but it is not limited to this. It can also be a geometric working window of other sizes, and different working windows can be selected according to the proportion of noise. In addition, image noise The attenuation effect and the image noise attenuation processing speed can adjust the number of image noise attenuation according to the user's needs. The relevant concept of the present invention was held in August 2006 in Beijing, China. International Symposium (2〇〇ICICIC, 2006 International Conference on Innovative Computing,

In Information and Control, "An Impulse" by Chao-Ho (Thou-Ho) Chen, Chao-Yu Chen, Tsong-Yi Chen, and Da-Jinn Wang Noise

Reduction Method by Adaptive Pixel-Correlation (a method of authentication for compressed data) and published (ρρ· 257-260). The above are only the preferred embodiments of the present invention, and all changes and modifications made to the scope of the invention are intended to be included in the scope of the present invention. [Simple description of the diagram] Figure 1 shows the second test image. Figure 2 is a graph of the total pixel correlation of the test image statistics in Figure 1. Figure 3 shows the test image with 20% random pulse noise added to the test image. 18 200826644 Figure 4 is a graph of the total pixel correlation of the test image statistics in Figure 3. Fig. 5 is a schematic diagram showing the proportion of the total pixel correlation after the random impulse noise is added to the test image of Fig. 1. Figure 6 is a schematic diagram of a working window according to an embodiment of the present invention. Figure 7 is a flow chart of the image noise attenuation method of the present invention. Fig. 8 is a view showing a comparison of the peak signal noise ratio of the prior art and the present invention by using the test image of Fig. 3 of Fig. 1. Figure 9 is a schematic view of the test image (a) of the present invention. Fig. 10 is a view showing a comparison of the peak signal noise ratio between the prior art and the present invention by using the test image of Fig. 1b. Fig. 1 is a schematic view showing another test image (b) of the present invention. [Description of main component symbols] 70 &gt; Flow 702 of pixel noise processing based on pixel correlation processing, calculating total pixel correlation of two adjacent thresholds (PCI.PC2) 704, calculating the total of two adjacent thresholds Pixel correlation difference (PCD) 706, determine the total pixel correlation difference (PCD) and get a total pixel correlation (PC1) 708, determine the total pixel correlation (PC1) and use a large window to attenuate image noise 710. Determine the total pixel correlation (PC1) and use a small window to attenuate the image noise 712, calculate a weighted similarity (Dh) between a center point pixel and a horizontal neighboring pixel, and a weighting similarity between a center point pixel and a vertical neighboring pixel. Sex (Dv) 19 200826644 714, judge Dh and Dv after the central point pixel direct output 71 6, judge Dh and Dv after the central point pixel is output by the median filter

Ti, the first preset value

Tc, second preset value T a, third preset value X0, center point pixel XI~X12, neighboring pixel of center point pixel 20

Claims (1)

200826644 X. Patent application scope: 1. A method for reducing image noise clothing, comprising the following steps: dividing an image into a geometric appearance working window, which is a central point pixel and adjacent pixels surrounding the central point pixel The work of the combination is to calculate the total pixel correlation of two adjacent thresholds; calculate the difference of the total pixel correlation of the two adjacent thresholds; determine whether the difference of the total pixel correlation is less than or equal to the first preset value, if less than or Equal to the first preset value to record a total pixel correlation; determine whether the total pixel correlation is less than the second preset value; less than the second preset value uses a large window attenuation to affect the noise; determine whether the total pixel correlation is Greater than or equal to the second preset value, if greater than or equal to the second preset value, a small window is used to attenuate the image noise; the weighted similarity between the center pixel and the horizontal neighbor pixel is calculated and a center point pixel and vertical Weighted similarity of neighboring pixels; determining weighted similarity between the center point pixel and the horizontal neighboring pixel and the center point pixel and the vertical neighboring pixel Whether the weight similarity is less than or equal to a third preset value, if the weighted similarity between the center point pixel and the horizontal neighboring pixel and the weighted similarity between the center point pixel and the vertical neighboring pixel are less than or equal to a third preset value, The center point pixel value is directly outputted; determining whether the weighted similarity between the center point pixel and the horizontal neighboring pixel and whether the weighted similarity between the center point pixel and the vertical neighboring pixel is greater than a third preset value, if the center point pixel is adjacent to the horizontal The weighted similarity of the pixels and the weighted similarity between the center point pixel and the vertical neighboring pixel are greater than the third preset 21 200826644 value, and the center point pixel value is rotated after the median filter. 2. The method for image noise reduction based on pixel concatenation according to the scope of the patent application, wherein the working window can be a diamond type, 3, for example, the scope of the patent application is ..... Image noise attenuation method, J: Based on the pixel correlation, the weighted value takes the distance of the Euclid distance: the pixel correlation is the data calculated from the reciprocal of the distance in each pixel. 4. An image processing system that operates using the method described in 4 or 2 or 3. 1 Use the patent scope 5, a signal processing 1 or the second or the third item: Ί use the method as claimed in the patent application. 6. A computer readable medium executing the code as claimed in the patent application, storing a code useful for a computer system attenuation method. &amp; Image 2 of the 2nd or 3rd item