CN103400351B - Low light based on KINECT depth map shines image enchancing method and system - Google Patents

Abstract

The invention discloses a low-light image enhancement method and system based on a KINECT depth map, comprising steps: step 1, collecting a depth map and a texture map of a low-light monitoring image through KINECT, and performing pre-processing on the depth map; step 2, Construct the depth information Gaussian kernel based on the depth values of the pixels in the depth map obtained in step 1 and its neighborhood pixels, and introduce the depth information Gaussian kernel as a kernel factor into the denoising algorithm based on the Gaussian kernel and denoise the texture map; steps 3. Obtain the normalized value of the depth information according to the depth information of the depth map obtained in step 1, introduce the normalized value of the depth information into the contrast stretching algorithm and perform contrast stretching on the denoised texture map. The present invention simultaneously utilizes the color information and the depth information of the image to realize low-light image enhancement, and further enhances the outline clarity, layering and visual effects of the low-light image.

Description

Low-light image enhancement method and system based on KINECT depth map

technical field

The invention relates to a low-light monitoring image enhancement method, in particular to a low-light image enhancement method and system based on a KINECT depth map.

Background technique

Currently, video surveillance systems are widely used in urban security systems. During the day, the image captured by the camera has better quality; however, at night or under low light conditions, due to the firmware characteristics of the CCD sensor, the captured image not only has more noise, but also has a narrow dynamic range of contrast. Visual enhancement of low-light surveillance images is critical if security systems are required to operate in low-light conditions.

The key technologies of low-light surveillance image enhancement include denoising and contrast stretching, which is undoubtedly a great challenge for images with low dynamic range and high noise. In 2007, Henrik ^[1] achieved image denoising by using anisotropy-based structured tensor filtering, and at the same time, used restricted histogram equalization to achieve image contrast enhancement. In 2010, QingXu ^[2] and others first denoised the image through the non-local mean algorithm in the time domain-space domain, then stretched the contrast through the tone mapping algorithm, and finally filtered in the YCbCr space; they also based on the motion information. The decision is made to improve the temporal filtering effect of the 3D non-local mean. In 2011, XuanDong ^[3] et al. reversed the low-light image, and then enhanced the reversed low-light image by using the dehazing algorithm. Although the above method has achieved certain results, there are still two problems as follows: (1) The above method only uses 2D information such as color and brightness in the image when enhancing low-light images, and does not use the depth information of the scene, resulting in The enhanced image lacks a sense of hierarchy; (2) The edge contour of the foreground target object is still not clear enough after enhancement.

The following references are involved in the article:

[1] H. Malm, M. Oskarsson et al, "Adaptive enhancement and noise reduction in very low light-level video," inProc.11thInternational Conference on Computer Vision, pp.1–8, 2007.

[2] QingXu, Hailin Jiang, Riccardo Scopigno, and Mateu Sbert, "A new approach for very dark video denoising and denhancement," in Proc.17thInternational Conference on Image Processing, pp.1185-1188, 2010

[3] XuanDong, GuanWang, Yi(Amy) Pangetal, "Fast efficient algorithm for enhancement of flowlighting video," inProc. IEEE International Conference on Multimedia and Expo, pp.1-6, 2011.

Contents of the invention

Aiming at the deficiencies in the prior art, the present invention proposes a low-light image enhancement method that utilizes the color information and depth information of the image at the same time, which can further enhance the outline definition, layering and visual effects of the low-light image.

In order to solve the problems of the technologies described above, the present invention adopts the following technical solutions:

1. A low-light image enhancement method based on the KINECT depth map, including steps:

Step 1, collect the depth map and texture map of the low-light monitoring image through KINECT, and pre-process the depth map;

Step 2: Construct the depth information Gaussian kernel based on the depth values of the pixels in the depth map obtained in step 1 and the depth values of their neighborhood pixels, introduce the depth information Gaussian kernel as a kernel factor into the denoising algorithm based on the Gaussian kernel, and de-noise the texture image noise;

Step 3, obtain the normalized value of the depth information according to the depth information of the depth map obtained in step 1, introduce the normalized value of the depth information into the contrast stretching algorithm to perform contrast stretching on the texture map after denoising, the described The normalized value of the depth information includes the normalized value of the depth value and the normalized value of the modulus of the gradient of the depth value.

Depth information Gaussian kernel as described in step 2 Among them, y represents the pixel in the depth map obtained in step 1; D(y) represents the depth value of pixel y; x is the neighborhood pixel of pixel y in the depth map obtained in step 1; D(x) is The depth value of the pixel x; h _d is the decay factor of the Gaussian kernel, and the value is taken according to the noise intensity in the texture map; ||·|| ₂ represents the two norm.

In step 3, introducing the normalized value of the depth information into the contrast stretching algorithm to perform contrast stretching on the denoised texture map includes performing global contrast stretching and local contrast stretching on the denoised texture map.

The introduction of the normalized value of the depth information into the contrast stretching algorithm and performing global contrast stretching on the denoised texture map further includes sub-steps:

3-1a Obtain the normalized value Norm(D(y)) of the depth value according to the depth value of the depth map obtained in step 1, and obtain the adaptive parameter b(y)=λ e ^-Norm(D(y)) , where λ is a constant, which is selected according to the degree of enhancement;

3-2a Introduce the adaptive parameter b(y) into the global contrast stretching algorithm to perform global contrast stretching on the denoised texture map.

The described method of introducing the normalized value of the depth information into the contrast stretching algorithm and performing local contrast stretching on the denoised texture map is as follows:

According to the normalized value of the modulus of the depth value gradient of the depth map obtained in step 1, the normalized value of the depth value or the normalized value of the modulus of the depth value gradient is introduced into the local contrast stretching algorithm for denoising Texture maps are locally contrast stretched.

2. Low-light image enhancement system based on KINECT depth map, including:

The pre-processing module is used to collect the depth map and texture map of the low-light monitoring image through KINECT, and pre-process the depth map;

The denoising module is used to construct the depth information Gaussian kernel based on the depth value of the pixel point in the depth map and its neighboring pixels, and introduce the depth information Gaussian kernel as a kernel factor into the denoising algorithm based on the Gaussian kernel and denoise the texture image ;

The contrast stretching module is used to obtain the normalized value of the depth information according to the depth information of the depth map, and introduces the normalized value of the depth information into the contrast stretching algorithm to perform contrast stretching on the texture map after denoising, the described The normalized value of the depth information includes the normalized value of the depth value and the normalized value of the modulus of the gradient of the depth value.

The above depth information Gaussian kernel Among them, y represents the pixel in the depth map; D(y) represents the depth value of pixel y; x is the neighborhood pixel of pixel y in the depth map; D(x) is the depth value of pixel x; h _d is the decay factor of the Gaussian kernel, and the value is taken according to the noise intensity in the texture map; ||·|| ₂ means the two norm.

The above-mentioned contrast stretching module further includes a global contrast stretching module and a local contrast stretching module, wherein:

The global contrast stretching module further includes submodules:

The adaptive parameter acquisition module is used to obtain the normalized value Norm(D(y)) of the depth value according to the depth value of the depth map, and obtain the adaptive parameter b(y)=λ·e ^{-Norm(D(y) )} , where λ is a constant, and the value is taken according to the degree of enhancement;

A stretching module is used to introduce the adaptive parameter b(y) into the global contrast stretching algorithm to perform global contrast stretching on the denoised texture map;

The local contrast stretching module is used to introduce the normalized value of the depth value or the normalized value of the modulus of the depth value gradient into the local contrast stretching algorithm according to the normalized value of the modulus of the depth value gradient of the depth map. The denoised texture map is subjected to local contrast stretching.

Compared with the prior art, the present invention has the following advantages:

The invention utilizes the depth information of the scene to enhance the low-light image, and the enhanced image outline is clearer, and has better layering and visual effects.

Description of drawings

Fig. 1 is a flow chart of the method of the present invention;

Figure 2 is the depth map and texture map of the low-light monitoring image collected by KINECT, where picture (a) is the depth map and picture (b) is the texture map;

Figure 3 is the depth map and texture map of another low-light monitoring image collected by KINECT, where picture (a) is a depth map and picture (b) is a texture map;

Fig. 4 is a comparison diagram of the enhancement effect of Fig. 2 enhanced by the present invention and the prior art, wherein, figure (a) is the enhancement effect of the present invention, and figure (b) is the enhancement effect of the prior art;

Fig. 5 is a comparison diagram of the enhancement effects of Fig. 3 enhanced by the present invention and the prior art, wherein, figure (a) is the enhancement effect of the present invention, and figure (b) is the enhancement effect of the prior art.

Detailed ways

The specific implementation of the present invention will be described in detail below.

The present invention is based on the low-light monitoring image enhancement method of KINECT depth map, comprises steps:

Step 1, collect the depth map and texture map of the low-light monitoring image through KINECT, and process the depth map to obtain a smooth depth map.

This step further includes sub-steps:

Step 1.1, collect the depth map and texture map of the low-light monitoring image through KINECT, and perform perspective registration on the depth map. In this specific implementation, view registration can be implemented using the OpenNI toolbox in PrimeSense.

Step 1.2: Hole filling and median filtering are performed on the depth map after viewing angle registration to obtain a smooth depth map. In this specific implementation, the hole filling is realized by using the in-painting algorithm, and the median filtering is realized by using the median filtering algorithm.

Step 2, based on the depth information of the depth map, the non-local mean denoising method is used to denoise the texture map.

In this step, the depth map is a smooth depth map obtained after processing in step 1; the texture map is a texture map of a low-light monitoring image collected by KINECT.

This step further includes sub-steps:

Step 2.1, build a Gaussian kernel based on the depth information of the depth map:

$e^{-\frac{{\left|\right|\mathrm{D.}\left(\mathrm{the\; y}\right)-\mathrm{D.}\left(x\right)\left|\right|}_2^2}{h_d^2}}$ Formula 1)

In formula (1):

y represents the pixel in the depth map;

D(y) represents the depth value of pixel point y in the depth map;

x is the neighborhood pixel of pixel y;

D(x) is the depth value of pixel x;

h _d is the decay factor of the Gaussian kernel, and the value is selected according to the noise intensity in the texture map. If the noise is large, the value is larger;

||·|| ₂ represents the second norm.

Based on the depth values of each pixel in the depth map and its neighboring pixels, formula (1) is used to construct the depth information Gaussian kernel corresponding to each pixel in the depth map.

In step 2.2, the depth information Gaussian kernel is added as a kernel factor to the non-local mean denoising algorithm, and the texture image is denoised. The specific denoising is realized based on formula (2):

$I^{\′}\left(\mathrm{the\; y}\right)=\underset{x\∈S_{\mathrm{the\; y}}}{\mathrm{\Σ}}\left(\frac{1}{Z\left(\mathrm{the\; y}\right)}{e}^{-\frac{{\left|\right|v\left(P_{\mathrm{the\; y}}\right)-v\left(P_x\right)\left|\right|}_{2,\mathrm{\α}}^2}{h^2}}{e}^{-\frac{{\left|\right|\mathrm{D.}\left(\mathrm{the\; y}\right)-\mathrm{D.}\left(x\right)\left|\right|}_2^2}{h_d^2}}I\left(x\right)\right)$ Formula (2)

In formula (2):

$Z\left(\mathrm{the\; y}\right)=\underset{x\∈S_{\mathrm{the\; y}}}{\mathrm{\Σ}}\left(e^{-\frac{{\left|\right|v\left(P_{\mathrm{the\; y}}\right)-v\left(P_x\right)\left|\right|}_{2,\mathrm{\α}}^2}{h^2}}{e}^{-\frac{{\left|\right|\mathrm{D.}\left(\mathrm{the\; y}\right)-\mathrm{D.}\left(x\right)\left|\right|}_2^2}{h_d^2}}\right),$ used to achieve normalization;

y is the current pixel point to be denoised in the texture image, P _y represents the image block centered on the current pixel point y to be denoised in the texture image, I'(y) is the current pixel point y to be denoised in the texture image After the pixel value;

x is the neighborhood pixel of the current pixel y to be denoised in the texture map, P _x represents the image block centered on the neighborhood pixel x in the texture map, and I(x) is the neighborhood pixel x in the texture map Pixel values;

S _y represents the neighborhood in the texture map centered on the current pixel point y to be denoised;

h is the decay factor of the pixel in the texture image, α is the standard deviation of the Gaussian kernel, h and α are valued according to the noise intensity in the texture image, the larger the noise, the larger the value of h and α;

||·|| _{2, α} means Gaussian weighted Euclidean distance;

v(P _x ) and v(P _y ) represent the column vectors composed of the pixel values of the pixels in the image blocks P _x and P _y respectively.

In this specific implementation, pixels at the same position in the depth map and the texture map are represented by the same letter.

Step 3, obtain the normalized value of the depth information according to the depth information of the depth map, introduce the normalized value of the depth information into the contrast stretching algorithm and perform contrast stretching on the denoised texture map.

The normalized value of the depth information in the present invention includes the normalized value of the depth value and the normalized value of the modulus of the gradient of the depth value.

In this specific implementation, performing contrast stretching on the texture map includes performing global contrast stretching and local contrast stretching on the texture map. In this step, the depth map is the smooth depth map obtained after processing in step 1.

This step further includes sub-steps:

In step 3.1, global contrast stretching is performed on the denoised texture map based on the depth information of the depth map, which can enhance the layering of the image while stretching the dynamic range of the image.

First, obtain the normalized depth value Norm(D(y)) according to the depth value of the depth map:

$\mathrm{Norm}\left(\mathrm{D.}\left(\mathrm{the\; y}\right)\right)=\frac{\mathrm{D.}\left(\mathrm{the\; y}\right)-{\mathrm{D.}}_{\min }}{{\mathrm{D.}}_{\max }-{\mathrm{D.}}_{\min }}$ Formula (3)

In formula (3):

y is the current pixel in the denoised texture map, and D(y) is the corresponding depth value of the current pixel y in the texture map in the depth map;

D _min and D _max are the minimum and maximum depth values of all pixels in the depth map, respectively.

Then, the adaptive parameter b(y) is obtained according to the normalized depth value Norm(D(y)):

b(y)=λ·e ^-Norm(D(y)) formula (4)

In formula (4), λ is a constant, and the value is selected according to the degree of enhancement. If you want to obtain a stronger degree of enhancement, you can take a smaller value; and vice versa.

Finally, based on the adaptive parameter b(y), the global contrast stretching is performed on the denoised texture image. In this specific implementation, the tone mapping enhancement algorithm is used to perform global contrast stretching on the texture image, as follows:

$I_{\mathrm{en}\_g}\left(\mathrm{the\; y}\right)=\log \left(\frac{I\left(\mathrm{the\; y}\right)+1}{I_{\max }+1}\right)\&Center\; Dot;\frac{\log 10}{\log (2+\left({\left(\frac{I\left(\mathrm{the\; y}\right)}{I_{\max }}\right)}^{\frac{\log b}{\log 0.5}}\right)\&Center\; Dot;8)}\·I_{\max }$ Formula (5)

In formula (5):

I(y) represents the pixel value of the current pixel point y in the R channel, G channel or B channel before global contrast stretching;

I _max is the maximum value of all pixels in the R channel, G channel or B channel;

I _{en_g} (y) is the pixel value corresponding to the current pixel point y after global contrast stretching.

The present invention introduces adaptive parameters into the global contrast stretching, so that the global contrast stretching can be adaptively adjusted according to the depth map information.

In this step, global contrast stretching is performed on the denoised texture image, specifically, formula (5) is executed on the R channel, G channel, and B channel of the texture image respectively.

In step 3.2, local contrast stretching is performed on the denoised texture map based on the depth information of the depth map to enhance the local layering of the image.

Because there is a strong depth change at the boundary between the foreground and the background of the texture image, the local contrast stretching is performed on the image to enhance the level difference between the foreground and the background. extension, as follows:

$I_{\mathrm{en}\_l}\left(\mathrm{the\; y}\right)=I_{\mathrm{en}\_g}\left(\mathrm{the\; y}\right)+\mathrm{Norm}\left(\mathrm{\α}\left(\mathrm{the\; y}\right)\right)\·(I_{\mathrm{en}\_g}\left(\mathrm{the\; y}\right)-{\stackrel{\‾}{I}}_{\mathrm{en}\_g}\left(\mathrm{the\; y}\right))$ Formula (6)

In formula (6):

I _{en_g} (y) is the pixel value corresponding to the current pixel point y after global contrast stretching;

I _{en_l} (x) is the pixel value corresponding to the current pixel point y after local contrast stretching;

is the pixel average value of the neighboring pixels of the current pixel point y;

α(y) represents the depth information weight discriminant:

$\mathrm{\α}\left(\mathrm{the\; y}\right)=\left\{\begin{array}{cc}\left|\right|\▿\mathrm{D.}\left(\mathrm{the\; y}\right)\left|\right|& \mathrm{if}\left|\right|\▿\mathrm{D.}\left(\mathrm{the\; y}\right)\left|\right|\≠0\\ \mathrm{D.}\left(\mathrm{the\; y}\right)& \mathrm{else}\end{array}\right.$ Formula (7)

Among them, ||▽D(y)|| represents the modulus of the gradient of the depth value D(y) of the pixel point y in the depth map.

When α(y)=||▽D(y)||, $\mathrm{Norm}\left(\mathrm{\α}\left(\mathrm{the\; y}\right)\right)=\frac{\left|\right|\▿\mathrm{D.}\left(\mathrm{the\; y}\right)\left|\right|-{\left|\right|\▿\mathrm{D.}\left|\right|}_{\min }}{{\left|\right|\▿\mathrm{D.}\left|\right|}_{\max }-{\left|\right|\▿\mathrm{D.}\left|\right|}_{\min }},$ Among them, ||▽D(y)|| represents the modulus of the gradient of the depth value D(y) of pixel y, and ||▽D|| _min and ||▽D|| _max are all pixels in the depth map, respectively The minimum and maximum values of the magnitude of the gradient of the depth value.

The technical solutions of the present invention will be further described below in conjunction with specific application examples.

First, the depth map and texture map of the low-light monitoring image are collected by KINECT, and the depth map is pre-processed; then, based on the non-local mean denoising algorithm and the tone mapping enhancement algorithm, the depth information of the depth map is added to achieve low-light Light image enhancement.

In this embodiment, the decay factor h of the pixels in the texture map and the decay factor h _d of the Gaussian kernel are both set to 15, the search range and image block size of the non-local mean algorithm are 21x21 and 7x7 respectively, and the standard deviation of the Gaussian kernel α= 0.15. Figure 2 and Figure 3 are the depth map and texture map of low-light surveillance images collected by KINECT, where Figure 2(a) and Figure 3(a) are depth maps, Figure 2(b) and Figure 3(b) are texture picture. Take Fig. 2～3 as example below, describe the method step of the present invention in detail:

Step 1, use the OpenNI toolbox in PrimeSense to perform perspective registration on the depth maps in Figure 2 and Figure 3, and then perform hole filling and filtering processing on the depth maps after perspective registration, specifically: use the in-painting algorithm The depth map is filled with holes, and the median filter algorithm is used to filter to obtain a smooth depth map.

Step 2, based on the depth map information obtained in step 1, denoise the texture map through a non-local mean denoising algorithm.

Step 2.1, for the pixel point y to be denoised in the texture map, within the search range of 21x21 centered on the pixel point y to be denoised in the depth map, calculate the center pixel point y and each neighboring pixel point within the search range Euclidean distance of x, and generate a Gaussian kernel Wherein: D(y) represents the depth value of pixel point y to be denoised in the depth map; D(x) is the depth value of neighborhood pixel point x; h _d is the decay factor of Gaussian kernel, in the present embodiment, h _d Take 15; ||·|| ₂ means two norms.

Step 2.2, add the Gaussian kernel of the depth map as a kernel factor to the non-local mean denoising algorithm to obtain the above formula (2) $I^{\′}\left(\mathrm{the\; y}\right)=\underset{x\∈S_{\mathrm{the\; y}}}{\mathrm{\Σ}}\left(\frac{1}{Z\left(\mathrm{the\; y}\right)}{e}^{-\frac{{\left|\right|v\left(P_{\mathrm{the\; y}}\right)-v\left(P_x\right)\left|\right|}_{2,\mathrm{\α}}^2}{h^2}}{e}^{-\frac{{\left|\right|\mathrm{D.}\left(\mathrm{the\; y}\right)-\mathrm{D.}\left(x\right)\left|\right|}_2^2}{h_d^2}}I\left(x\right)\right),$ The texture map is denoised using formula (2), where: $Z\left(\mathrm{the\; y}\right)=\underset{x\∈S_{\mathrm{the\; y}}}{\mathrm{\Σ}}\left(e^{-\frac{{\left|\right|v\left(P_{\mathrm{the\; y}}\right)-v\left(P_x\right)\left|\right|}_{2,\mathrm{\α}}^2}{h^2}}{e}^{-\frac{{\left|\right|\mathrm{D.}\left(\mathrm{the\; y}\right)-\mathrm{D.}\left(x\right)\left|\right|}_2^2}{h_d^2}}\right)$ is the normalization process; P _y represents the image block centered on the current pixel point y to be denoised in the texture image, and the size of the image block is 7x7; I'(y) is the current pixel point y to be denoised in the texture image. The pixel value after noise; P _x represents the image block centered on the neighborhood pixel point x in the texture map, and the size of the image block is 7x7; I(x) is the pixel value of the neighborhood pixel point x in the texture map;|| ·|| _{2, α} represents the Gaussian weighted Euclidean distance; v(P _x ) and v(P _y ) represent the column vectors composed of the pixel values of each pixel in the image block P _x and P _y respectively; S _y represents the texture map The neighborhood centered on the current pixel point y to be denoised, the size of the neighborhood is the search range of the non-local mean algorithm, which is 21x21; h is the decay factor of the pixel in the texture map, and α is the standard deviation of the Gaussian kernel. This implementation In the example, h is set to 15 and α is set to 0.15.

Step 3, use the depth information of the depth map to perform contrast stretching on the denoised texture map.

The contrast stretching in the present invention includes global contrast stretching and local contrast stretching.

Step 3.1, using the depth information of the depth map to perform global contrast stretching on the denoised texture map.

First, calculate the normalized depth value Norm(D(y)) based on the depth information of the depth map;

Then, based on the normalized depth value Norm(D(y)), the adaptive parameter b(y)=λ·e ^{-Norm(D(y)) is} obtained, where λ is a constant, and in this embodiment, λ is set is 0.6.

Finally, based on the adaptive parameter b(y), the above formula (5) is used to perform global contrast stretching on the denoised texture map.

Step 3.2, local contrast stretching of the denoised texture map based on the depth information of the depth map.

Due to the strong depth variation at the boundary of the image foreground and background, local contrast stretching is performed on the image to enhance the layer difference of the foreground and background.

This specific implementation is based on the depth information of the depth map, and uses the formula (6) to perform local contrast stretching on the denoised texture map.

In this embodiment, Fig. 2 and Fig. 3 are enhanced, and the enhancement effects are shown in Fig. 4(a) and Fig. 5(a) respectively. At the same time, in this specific implementation, the traditional non-local mean denoising algorithm and tone mapping enhancement algorithm are used to denoise and contrast stretch Figure 2 and Figure 3, and the enhancement effects are shown in Figure 4(b) and Figure 5(b). It can be seen from Figures 4 to 5 that the outline of the enhanced image obtained by the present invention is clearer, and has better layering and visual effects.

The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which the present invention belongs can make various modifications or supplements to the described specific embodiments or adopt similar methods to replace them, but they will not deviate from the spirit of the present invention or go beyond the definition of the appended claims range.

Claims (2)

1. based on the low-light image enhancement method of KINECT depth map, it is characterized in that, comprising steps: Step 1, collect the depth map and texture map of the low-light surveillance image through KINECT, and perform perspective registration, hole filling, and median filtering on the depth map; Step 2, construct depth information Gaussian kernel based on the depth values of the pixels in the depth map obtained in step 1 and their neighboring pixels Introduce the depth information Gaussian kernel as a kernel factor into the denoising algorithm based on the Gaussian kernel and denoise the texture image, where y represents the pixel in the depth map obtained in step 1; D(y) represents the depth value of pixel y ; x is the neighborhood pixel of pixel y in the depth map obtained in step 1; D(x) is the depth value of pixel x; h _d is the decay factor of the Gaussian kernel, and the value is taken according to the noise intensity in the texture map;| |·|| ₂ means two norms; Step 3, obtain the normalized value of the depth information according to the depth information of the depth map obtained in step 1, introduce the normalized value of the depth information into the contrast stretching algorithm to perform contrast stretching on the texture map after denoising, the described The normalized value of the depth information includes the normalized value of the depth value and the normalized value of the modulus of the depth value gradient; the described normalized value of the depth information is introduced into the contrast stretching algorithm to the denoised texture map Performing contrast stretching includes performing global contrast stretching and local contrast stretching on the denoised texture map, where: Introducing the normalized value of the depth information into the contrast stretching algorithm to perform global contrast stretching on the denoised texture map further includes sub-steps: 3-1a Obtain the normalized value Norm(D(y)) of the depth value according to the depth value of the depth map obtained in step 1, and obtain the adaptive parameter b(y)=λ e ^-Norm(D(y)) , where λ is a constant, which is selected according to the degree of enhancement; 3-2a Introduce the adaptive parameter b(y) into the global contrast stretching algorithm to perform global contrast stretching on the denoised texture map; The normalized value of the depth information is introduced into the contrast stretching algorithm to perform local contrast stretching on the denoised texture map as follows: According to the normalized value of the modulus of the depth value gradient of the depth map obtained in step 1, the normalized value of the depth value or the normalized value of the modulus of the depth value gradient is introduced into the local contrast stretching algorithm for denoising Texture maps are locally contrast stretched. 2. The low-light image enhancement system based on the KINECT depth map is characterized in that it comprises: The pre-processing module is used to collect the depth map and texture map of the low-light monitoring image through KINECT, and perform viewing angle registration, hole filling, and median filtering on the depth map; The denoising module is used to construct the depth information Gaussian kernel based on the depth values of the pixels in the depth map and their neighboring pixels Introduce the depth information Gaussian kernel as a kernel factor into the denoising algorithm based on the Gaussian kernel and denoise the texture image, where y represents the pixel in the depth map obtained in step 1; D(y) represents the depth value of pixel y ; x is the neighborhood pixel of pixel y in the depth map obtained in step 1; D(x) is the depth value of pixel x; h _d is the decay factor of the Gaussian kernel, and the value is taken according to the noise intensity in the texture map;| |·|| ₂ means two norms; The contrast stretching module is used to obtain the normalized value of the depth information according to the depth information of the depth map, and introduces the normalized value of the depth information into the contrast stretching algorithm to perform contrast stretching on the denoised texture map, the described The normalized value of the depth information includes the normalized value of the depth value and the normalized value of the modulus of the depth value gradient; The contrast stretching module further includes a global contrast stretching module and a local contrast stretching module, wherein the global contrast stretching module further includes submodules: The adaptive parameter acquisition module is used to obtain the normalized value Norm(D(y)) of the depth value according to the depth value of the depth map, and obtain the adaptive parameter b(y)=λ·e ^{-Norm(D(y) )} , where λ is a constant, and the value is taken according to the degree of enhancement; A stretching module is used to introduce the adaptive parameter b(y) into the global contrast stretching algorithm to perform global contrast stretching on the denoised texture map; The local contrast stretching module is used to introduce the normalized value of the depth value or the normalized value of the modulus of the depth value gradient into the local contrast stretching algorithm according to the normalized value of the modulus of the depth value gradient of the depth map. The denoised texture map is subjected to local contrast stretching.