CN104463813B - Infrared image noise reduction method based on noise recognition - Google Patents

Abstract

The invention discloses an infrared image noise reduction method based on noise recognition. The method introduces the basic idea of noise recognition, respectively calculates the membership degree of the current pixel based on the censored mean value and gradient based, and investigates the influence of the current pixel on noise interference. The joint criterion is used to judge whether the current pixel is a noise pixel, and finally the noise reduction is performed according to the judgment result to realize the noise reduction of the infrared image. The invention has a small amount of calculation and is easy to implement in real time; compared with traditional algorithms, it can more effectively protect image edges and details; in the process of denoising noise points, the texture gradient information of the image is also considered, and the original signal is more accurately Make an estimate.

Description

A Noise Reduction Method for Infrared Image Based on Noise Recognition

technical field

The invention belongs to the technical field of infrared image processing, and in particular relates to an infrared image noise reduction method based on noise recognition.

Background technique

The infrared imaging system has strong anti-interference ability, good concealment performance, strong atmospheric penetration ability, and is suitable for a variety of special occasions. It has more and more applications in scientific research, military, medical, industrial, and civil applications. However, due to factors such as the production process, sensitivity of the infrared detector, and the radiation characteristics of the target and the environment, the contrast of the infrared thermal image is not high compared with the visible light image, showing the characteristics of high background, low contrast, and obvious noise, which is not conducive to later use. In order to make full use of captured information, suppress noise, improve image quality, and facilitate higher-level processing, noise reduction must be performed on infrared images.

Traditional image denoising methods are mainly divided into three categories: temporal denoising, spatial denoising, and frequency domain denoising. Time-domain noise reduction makes use of the signal acquisition process, the signal has a strong correlation, and the noise has the characteristics of random distribution, the signal of the same pixel between frames is averaged to obtain the effect of noise reduction, but in the scene of high-speed motion It will cause image blur and smear; spatial noise reduction is to use the spatial correlation of adjacent pixels to perform noise reduction. Typical methods are mean filtering, median filtering, Wiener filtering, etc. The algorithm is simple to implement and the operation speed is fast. , the disadvantage is that the image will be blurred while reducing the noise, especially at the edge and details of the object; the frequency domain noise reduction is to transform the image from the spatial domain to the frequency domain through image transformation, and use the filtering method to filter out the high frequency part representing the noise , but some noise whose frequency components are similar to the signal cannot be removed, and poor selection of the filtering threshold has a great impact on the noise reduction effect. In addition, there are some methods that combine the above noise reduction principles to denoise images from multiple aspects. For example, wavelet denoising combines the principles of spatial domain and frequency domain denoising, and has good localization properties and multi-scale analysis. It can effectively separate the signal from the noise, but it has a large amount of computation.

With the development of infrared imaging system, the imaging resolution of the system is getting higher and higher, which makes more and more image data need to be processed in real time. Due to limited system resources, some noise reduction algorithms that require a lot of calculation and occupy a lot of storage resources are not applicable. In order to realize the real-time processing of infrared images, it is necessary to study an infrared image denoising algorithm which has a small amount of calculation and is easy to implement in real time.

Contents of the invention

In order to solve the above-mentioned technical problems, the present invention provides a noise-recognition-based infrared image noise reduction method.

The technical scheme adopted in the present invention is: a noise-recognition-based infrared image noise reduction method, characterized in that it comprises the following steps:

Step 1: Calculate the membership degree of the current pixel based on the truncated mean value, including the noise membership degree μ _n (i,j) based on the truncated mean value and the signal membership degree μ _s (i,j) based on the truncated mean value, where i and j is the coordinate of the current pixel;

Step 2: Calculate the gradient-based membership of the current pixel, including the gradient-based noise membership S _n (i,j) and the gradient-based signal membership S _s (i,j);

Step 3: According to the membership degree based on the truncated mean value and the membership degree based on the gradient calculated in step 1 and step 2, determine whether the current pixel is a noise pixel;

Step 4: If the current pixel is a noise pixel, perform noise reduction processing on the pixel, and return to step 1 until the entire image is traversed after processing; if the current pixel is a normal signal pixel, directly return to step 1 until the entire image is traversed image.

As a preference, the calculation formula of the noise membership μ _n (i, j) based on the censored mean value described in step 1 is:

Among them, f(i,j) is the pixel gray value of coordinates (i,j), a and b are variable parameters, and empirical values are obtained according to experiments, T(i,j) is a 3× pixel centered on the current pixel The censored mean of the 3 windows, the calculation formula is:

Among them, A _i,j represents the set of all pixel gray levels in the 3×3 window centered on the current pixel, pMax is the maximum gray level in the set A _{i,j ,} and pMin is the minimum gray level in the set A i,j ;

The calculation formula of the signal membership μ _s (i,j) based on the censored mean in step 1 is:

μ _s (i,j)=1−μ _n (i,j).

As a preference, the calculation formula of the gradient-based noise membership degree S _n (i, j) described in step 2 is:

Among them, d represents the direction, and there are 8 directions in total, which are upper U, upper left LU, upper right RU, left L, right R, lower left LD, lower D, and lower right RD. F _n ^d (i, j) is the coordinate (i , j) based on the noise membership of the gradient in the d direction, the calculation formula is:

in:

When d=U, LU, RU, L, R, LD, D, RD, Respectively equal to |f(i–1,j)–f(i,j)|, |f(i–1,j–1)–f(i,j)|, |f(i–1,j+1 )–f(i,j)|, |f(i,j–1)–f(i,j)|, |f(i,j+1)–f(i,j)|, |f(i +1,j–1)–f(i,j)|, |f(i+1,j)–f(i,j)|, |f(i+1,j+1)–f(i, j)|;

When d=U, LU, RU, L, R, LD, D, RD, Respectively equal to |f(i–1,j–1)–f(i,j–1)|, |f(i,j–1)–f(i+1,j)|, |f(i–1 ,j)–f(i,j–1)|, |f(i+1,j–1)–f(i+1,j)|, |f(i–1,j+1)–f( i–1,j)|, |f(i+1,j)–f(i,j+1)|, |f(i+1,j+1)–f(i,j+1)|, |f(i,j+1)–f(i–1,j)|;

When d=U, LU, RU, L, R, LD, D, RD, Respectively equal to |f(i–1,j+1)–f(i,j+1)|, |f(i–1,j–1)–f(i,j)|, |f(i,j +1)–f(i+1,j)|, |f(i–1,j–1)–f(i–1,j)|, |f(i+1,j+1)–f( i+1,j)|, |f(i,j–1)–f(i–1,j)|, |f(i+1,j–1)–f(i,j–1)|, |f(i+1,j)–f(i,j–1)|;

The function β( ) is defined as follows:

Among them, c and d are variable parameters, and empirical values are obtained according to experiments;

The calculation formula of the gradient-based signal membership degree S _s (i,j) described in step 2 is:

Among them, F _s ^d (i,j) is the signal membership degree of the current pixel based on the gradient in the d direction, and the calculation formula is:

Preferably, the method for judging whether the current pixel is a noise pixel in step 3 is: when μ _n (i,j)·S _n (i,j) is greater than or equal to μ _s (i,j)·S _s ( i,j), the current pixel is determined to be a noise pixel; when μ _n (i,j) S _n (i,j) is less than μ _s (i,j) S _s (i,j), the current pixel is determined is a normal signal pixel.

As a preference, the noise reduction processing for the pixel described in step 4 is implemented by searching for the order The smallest direction, denote this direction as d _min , then the gray level of the current pixel is replaced by the gray level calculated by the following formula:

This method introduces the basic idea of noise recognition, calculates the membership degree of the current pixel based on the censored mean value and gradient based respectively, examines the degree of noise interference of the current pixel, and uses the joint criterion to judge whether the current pixel is a noise pixel, and finally according to The judgment result is denoised to realize the denoising of the infrared image. The present invention has the following advantages:

1. The amount of calculation is small, and it is easy to implement in real time. Since the algorithm only performs statistical calculations on a single pixel and its 8 neighbors, the algorithm has low complexity and does not need to occupy a large amount of storage resources for caching frequency domain data, so the amount of calculation is small and it is easy to implement in real time;

2. Compared with traditional algorithms, it can protect image edges and details more effectively. Since the degree of noise interference of the current pixel is first investigated in the noise reduction process, the joint criterion is used to judge whether the current pixel is a noise pixel, which improves the scientificity and accuracy of noise judgment and avoids unnecessary noise reduction for non-noise points The image blur introduced after;

3. In the process of denoising the noise point, the texture gradient information of the image is also considered, and the original signal is estimated more accurately.

Description of drawings

Fig. 1: is the method flowchart of the present invention.

Fig. 2: is an original infrared image with a resolution of 512×640 of the embodiment of the present invention.

Fig. 3 is an infrared image artificially added with random noise on the original infrared image with a resolution of 512×640 according to an embodiment of the present invention.

Fig. 4 is a schematic diagram of a 3×3 window and eight directions according to an embodiment of the present invention.

Fig. 5: is the image processed by this method according to the embodiment of the present invention.

Fig. 6: is the image processed by the classical median filtering method according to the embodiment of the present invention.

detailed description

In order to facilitate those of ordinary skill in the art to understand and implement the present invention, the present invention will be described in further detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the implementation examples described here are only used to illustrate and explain the present invention, and are not intended to limit this invention.

Please see Fig. 1, the present invention mainly consists of 4 steps: calculating the membership degree of the current pixel based on the censored mean, calculating the membership degree of the current pixel based on the gradient, judging whether the current pixel is a noise pixel according to the calculated membership degree, if judging If the current pixel is a noise pixel, noise reduction processing is performed on the pixel.

Please refer to FIG. 2 and FIG. 3 , which are an original infrared image with a resolution of 512×640 according to the embodiment of the present invention. And the infrared image after artificially adding random noise on the original infrared image with a resolution of 512×640. Taking the original infrared image as an example below, each step of the present invention is described in detail:

Step (1) starts to traverse the entire image with a 3×3 window. Calculate the membership degree of the center pixel of the current window based on the truncated mean value, including the noise membership degree μ _n (i,j) based on the truncated mean value and the signal membership degree μ _s (i,j) based on the truncated mean value, where i and j is the coordinates where the current pixel is located, as shown in Figure 4;

The noise membership μ _n (i,j) based on the censored mean is calculated by the following formula:

Among them, f(i,j) is the pixel gray value of coordinates (i,j), a and b are variable parameters, and the empirical values are taken according to the experiment. If the system noise level is low, in order to preserve the details as much as possible, the noise The judgment should be loose, so a and b can take a large value (greater than 1% of the maximum gray scale). In this example, a and b are the empirical values of 20 and 80 respectively according to the experiment (the maximum gray scale is 16384), and T( i, j) is the truncated mean value of the 3×3 window centered on the current pixel, and the calculation formula is as follows:

Among them, A _i,j represents the set of all pixel gray levels in the 3×3 window centered on the current pixel, pMax is the maximum gray level in the set A _{i,j ,} and pMin is the minimum gray level in the set A i,j ;

The signal membership μ _s (i,j) based on the censored mean is calculated by the following formula:

μ _s (i, j) = 1-μ _n (i, j);

Step (2) Calculate the gradient-based membership of the current pixel, including the gradient-based noise membership S _n (i, j) and the gradient-based signal membership S _s (i, j);

Gradient-based noise membership S _n (i, j) is calculated using the following formula:

Among them, d represents the direction, a total of 8 directions, namely upper U, upper left LU, upper right RU, left L, right R, lower left LD, lower D, lower right RD, F _n ^d (i, j) is the current pixel based on The noise membership degree of the gradient in the d direction, the calculation formula is as follows:

Wherein, when d=U, LU, RU, L, R, LD, D, RD, Respectively equal to |f(i–1,j)–f(i,j)|, |f(i–1,j–1)–f(i,j)|, |f(i–1,j+1 )–f(i,j)|, |f(i,j–1)–f(i,j)|, |f(i,j+1)–f(i,j)|, |f(i +1,j–1)–f(i,j)|, |f(i+1,j)–f(i,j)|, |f(i+1,j+1)–f(i, j)|; When d=U, LU, RU, L, R, LD, D, RD, Respectively equal to |f(i–1,j–1)–f(i,j–1)|, |f(i,j–1)–f(i+1,j)|, |f(i–1 ,j)–f(i,j–1)|, |f(i+1,j–1)–f(i+1,j)|, |f(i–1,j+1)–f( i–1,j)|, |f(i+1,j)–f(i,j+1)|, |f(i+1,j+1)–f(i,j+1)|, |f(i,j+1)–f(i–1,j)|; when d=U, LU, RU, L, R, LD, D, RD, Respectively equal to |f(i–1,j+1)–f(i,j+1)|, |f(i–1,j–1)–f(i,j)|, |f(i,j +1)–f(i+1,j)|, |f(i–1,j–1)–f(i–1,j)|, |f(i+1,j+1)–f( i+1,j)|, |f(i,j–1)–f(i–1,j)|, |f(i+1,j–1)–f(i,j–1)|, |f(i+1,j)–f(i,j–1)|, the function β( ) is defined as follows:

Among them, c and d are variable parameters. According to the empirical value obtained from the experiment, if the system noise level is low, in order to preserve the details as much as possible, the noise judgment should be loose, so c and d can be appropriately large values (greater than the maximum gray level 0.5%), c, d get experience value 20,55 according to experiment respectively in this example;

Gradient-based signal membership S _s (i,j) is calculated using the following formula:

Among them, F _s ^d (i,j) is the signal membership degree of the current pixel based on the gradient in the d direction, and the calculation formula is as follows:

Step (3) judges whether the current pixel is a noise pixel according to the degree of membership calculated in step (1) and step (2);

According to the four membership degrees calculated in step (1) and step (2), when μ _n (i,j) S _n (i,j) is greater than or equal to μ _s (i,j) S _s (i,j ), it is determined that the current pixel is a noise pixel; when μ _n (i,j) S _n (i,j) is less than μ _s (i,j) S _s (i,j), it is determined that the current pixel is a normal signal pixel;

Step (4) If step (3) judges that the current pixel is a noise pixel, then perform noise reduction processing on the pixel, and return to step (1) after processing until the entire image is traversed; if step (3) judges that the current pixel is normal Signal pixels, then go back to step (1) until the entire image is traversed. The specific noise reduction methods are as follows:

Find the order in the 8 directions of U, LU, RU, L, R, LD, D, and RD The smallest direction, denote this direction as d _min , then the gray level of the current pixel is replaced by the gray level calculated by the following formula:

Please see accompanying drawing 5, which is an image processed by the present invention; please refer to accompanying drawing 6, which is an image processed by a classic median filtering algorithm. Comparing Figure 2 and Figure 3, we can see that Figure 3 adds a lot of noise compared to Figure 2, and the image quality is seriously degraded; comparing Figure 2, Figure 5 and Figure 6, we can see that only a small part of noise remains in Figure 5, and the image is basically It is restored to be close to that in Figure 2, but in Figure 6, it can still be clearly seen that there are many residual noises, the image details are blurred, and the noise reduction effect is not as good as that in Figure 5.

It should be understood that the parts not described in detail in this specification belong to the prior art.

It should be understood that the above-mentioned descriptions for the preferred embodiments are relatively detailed, and should not therefore be considered as limiting the scope of the patent protection of the present invention. Within the scope of protection, replacements or modifications can also be made, all of which fall within the protection scope of the present invention, and the scope of protection of the present invention should be based on the appended claims.

Claims (3)

1. a kind of infrared image noise-reduction method based on Noise Identification is it is characterised in that comprise the following steps：

Step 1：Calculate the degree of membership based on trimmed mean for the current pixel, including noise degree of membership μ based on trimmed mean_n(i,j) And signal degree of membership μ based on trimmed mean_s(i, j), wherein i and j are current pixel place coordinate；

Wherein said noise degree of membership μ based on trimmed mean_nThe computing formula of (i, j) is：

Wherein, f (i, j) is the grey scale pixel value of coordinate (i, j), and a, b are variable element, take empirical value, T (i, j) according to experiment It is the trimmed mean of 3 × 3 windows centered on current pixel, computing formula is：

Wherein, A_i,jRepresent the set of all pixels gray scale in 3 × 3 windows centered on current pixel, pMax is set A_i,j In maximum gray scale, pMin be set A_i,jIn minimal gray；

Described signal degree of membership μ based on trimmed mean_sThe computing formula of (i, j) is：

μ_s(i, j)=1- μ_n(i,j)；

Step 2：Calculate the degree of membership based on gradient for the current pixel, including noise degree of membership S based on gradient_n(i, j) and it is based on Signal degree of membership S of gradient_s(i,j)；

Wherein said noise degree of membership S based on gradient_nThe computing formula of (i, j) is：

Wherein, d represents direction, totally 8 directions, is upper U, upper left LU, upper right RU, left L, right R, lower-left LD, lower D, bottom right respectively RD,For the noise degree of membership based on gradient on d direction for the pixel of coordinate (i, j), computing formula is：

Wherein：

As d=U, LU, RU, L, R, LD, D, RD,Respectively equal to | f (i 1, j) f (i, j) |, | f (i 1, j 1) f (i, j) |、|f(i–1,j+1)–f(i,j)|、|f(i,j–1)–f(i,j)|、|f(i,j+1)–f(i,j)|、|f(i+1,j–1)–f(i,j) |、|f(i+1,j)–f(i,j)|、|f(i+1,j+1)–f(i,j)|；

As d=U, LU, RU, L, R, LD, D, RD,Respectively equal to | f (i 1, j 1) f (i, j 1) |, | f (i, j 1) f (i+ 1,j)|、|f(i–1,j)–f(i,j–1)|、|f(i+1,j–1)–f(i+1,j)|、|f(i–1,j+1)–f(i–1,j)|、|f(i+1, j)–f(i,j+1)|、|f(i+1,j+1)–f(i,j+1)|、|f(i,j+1)–f(i–1,j)|；

As d=U, LU, RU, L, R, LD, D, RD,Respectively equal to | f (i 1, j+1) f (i, j+1) |, | f (i 1, j 1) f (i,j)|、|f(i,j+1)–f(i+1,j)|、|f(i–1,j–1)–f(i–1,j)|、|f(i+1,j+1)–f(i+1,j)|、|f(i, j–1)–f(i–1,j)|、|f(i+1,j–1)–f(i,j–1)|、|f(i+1,j)–f(i,j–1)|；

Function β () is defined as follows：

Wherein, c, d are variable element, take empirical value according to experiment；

Described signal degree of membership S based on gradient_sThe computing formula of (i, j) is：

Wherein, F_s ^d(i, j) is the signal degree of membership based on gradient on d direction for the current pixel, and computing formula is：

Step 3：According to step 1, the calculated degree of membership based on trimmed mean of step 2 and the degree of membership based on gradient, sentence Whether disconnected current pixel is noise pixel；

Step 4：If current pixel be noise pixel, noise reduction process is carried out to this pixel, return to after having processed step 1 until Travel through entire image；If current pixel is normal signal pixel, it is returned directly to step 1 until having traveled through entire image.

2. the infrared image noise-reduction method based on Noise Identification according to claim 1 it is characterised in that：Institute in step 3 That states judges whether current pixel is noise pixel, and its method is：Work as μ_n(i,j)·S_n(i, j) is more than or equal to μ_s(i,j)·S_s When (i, j), judge current pixel as noise pixel；Work as μ_n(i,j)·S_n(i, j) is less than μ_s(i,j)·S_sWhen (i, j), judge to work as Preceding pixel is normal signal pixel.

3. the infrared image noise-reduction method based on Noise Identification according to claim 1 it is characterised in that：Institute in step 4 That states carries out noise reduction process to this pixel, and its concrete methods of realizing is to find in this 8 directions of U, LU, RU, L, R, LD, D, RD OrderMinimum direction, remembers that this direction is d_min, then the gray scale that current pixel gray scale is calculated with equation below Replace：