Deep learning denoising method for images in low-light scenes at night and in tunnels

Abstract

The present invention belongs to the field of image processing technology, and specifically relates to a deep learning denoising method for images in low-light scenes at night and in tunnels. The method comprises: step 1: nonlinearly mapping the target low-light scene image, and then adaptively enhancing it to obtain an intermediate image; step 2: multi-scale decomposition of the intermediate image using wavelet transform, and estimating the noise standard deviation at each scale; step 3: calculating the adaptive threshold of each scale and spatial position based on the estimated noise standard deviation and ambient light intensity; considering the tunnel effect, compensating the tunnel entrance area and tunnel exit area corresponding to the sharp change in brightness in the intermediate image to obtain a compensated image, and applying the adaptive threshold of each scale spatial position to the compensated image to obtain a denoised image. The present invention significantly improves the denoising effect and detail retention capability of low-light images, and enhances the adaptability of the method.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and particularly relates to a deep learning denoising method for images in low-light scenes at night and in tunnels. Background Art

[0002] In the field of modern image processing, especially in low-light scenes such as nighttime and tunnels, image noise reduction and enhancement technology has always been a hot topic and difficulty in research. Under low-light conditions, due to insufficient light, the number of photons received by the image sensor is small, resulting in a significant decrease in the image signal-to-noise ratio (SNR), which in turn generates a large amount of noise. These noises not only reduce the clarity of the image, but also cover up important detail information, affecting the effect of subsequent image analysis and processing. Therefore, image noise reduction and enhancement technology for low-light scenes has important research value and application prospects.

[0003] At present, a variety of technologies and methods have been proposed to solve the problem of image denoising in low-light scenes. The main methods include but are not limited to the following: Spatial domain denoising methods directly process image pixels. Common methods include mean filtering, median filtering and adaptive filtering. Mean filtering smoothes the image by taking the average value of neighboring pixels to reduce noise, but it also blurs image details. Median filtering replaces the central pixel value with the median of neighboring pixels, which can effectively remove salt and pepper noise, but has poor effect on Gaussian noise. Adaptive filtering adjusts the filtering parameters according to the local statistical characteristics of the image, which can retain image details to a certain extent, but its effect depends on the selection of parameters and the accuracy of local statistical characteristics. Frequency domain denoising methods process noise in the frequency domain by performing frequency domain transforms (such as Fourier transform, wavelet transform, etc.) on the image. Wavelet transform is a commonly used multi-scale analysis tool. By decomposing the image into different scales and frequency components, it can more accurately identify and remove noise. Wavelet threshold denoising method is a typical representative of frequency domain denoising. By setting an appropriate threshold, the noise in the high-frequency part is removed while retaining the image details in the low-frequency part. However, how to choose the appropriate threshold and wavelet basis function is still a challenge. The statistical model-based method performs noise reduction by establishing a statistical model of the image and noise. For example, the Gauss-Markov random field (GMRF) model, Bayesian estimation and other methods remove noise by probabilistically modeling the image and using techniques such as maximum a posteriori estimation (MAP). This type of method can better preserve image details, but has high computational complexity and is highly dependent on model assumptions. Summary of the invention

[0004] In view of this, the main purpose of the present invention is to provide a deep learning denoising method for images in low-light scenes at night and in tunnels. The present invention significantly improves the denoising effect and detail retention capability of low-light images and enhances the adaptability of the method.

[0005] The technical solution adopted by the present invention is as follows:

[0006] A deep learning denoising method for images in low-light scenes at night and in tunnels, the method comprising:

[0007] Step 1: After nonlinear mapping of the target low-light scene image, adaptive enhancement is performed to obtain an intermediate image;

[0008] Step 2: Use wavelet transform to decompose the intermediate image into multiple scales and estimate the noise standard deviation at each scale;

[0009] Step 3: Calculate the adaptive threshold at each scale and spatial position based on the estimated noise standard deviation and ambient light intensity; Consider the tunnel effect, compensate the tunnel entrance area and tunnel exit area corresponding to the sharp change in brightness in the intermediate image to obtain a compensated image, and apply the adaptive threshold at each scale spatial position to the compensated image to obtain a denoised image.

[0010] Furthermore, the low-light scene image is an image acquired in an environment where the light intensity is less than or equal to 100 lux.

[0011] Furthermore, the target low-light scene image is given by the following formula: Perform nonlinear mapping to obtain a nonlinear mapping image :

[0012] ;

[0013] in, is the target low-light scene image, with a value range of [0,1]; for The global average brightness of is the error function, and its specific form is as follows:

[0014] ;

[0015] in, is the independent variable of the error function; is the integration variable.

[0016] Furthermore, in step 1, the nonlinear mapping image Perform adaptive enhancement to obtain the intermediate image The formula is as follows:

[0017] ;

[0018] in, yes The global pixel mean of ; yes The global pixel standard deviation of for The global average brightness of and They are The width and height of and The horizontal and vertical coordinates of the pixels are ； is an adaptive parameter, and its calculation formula is as follows:

[0019] ;

[0020] is the gamma correction parameter, calculated as follows:

[0021] .

[0022] Furthermore, in step 2, after multi-scale decomposition of the intermediate image using wavelet transform, the noise standard deviation is estimated at each scale using the following formula:

[0023] ;

[0024] in, For the Scale at light level The standard deviation of noise under For the The scale of wavelet coefficients; For the The mean of the wavelet coefficients at each scale; For the The number of wavelet coefficients at each scale; is the set adjustment parameter; is the central scale; For the The maximum value of the wavelet coefficients at each scale.

[0025] Furthermore, the adaptive threshold is calculated for each scale and spatial position by the following formula:

[0026] ;

[0027] in, For the At this scale, pixel position Adaptive threshold of is the reference light intensity, equal to 100 lux; pixel position , is the horizontal axis, is the vertical coordinate; the center pixel position , is the horizontal coordinate of the center pixel position; is the ordinate of the center pixel position; , , and are control parameters, among which, is an integer ranging from 2 to 4; The value range is from 0 to 1; The value range of is 0.1 to 0.3; ; To find the determinant of a matrix.

[0028] Furthermore, in step 3, the tunnel effect is taken into account, and the tunnel entrance area and tunnel exit area corresponding to the sharp change in brightness in the intermediate image are compensated by the following formula to obtain a compensated image: :

[0029] ;

[0030] in, is the preset tunnel effect compensation function, which is calculated using the following formula:

[0031] ;

[0032] in, To compensate for the strength; are the coordinates of the tunnel entrance center, is the horizontal axis, is the vertical coordinate; is the center coordinate of the tunnel exit, is the horizontal axis, is the vertical coordinate; is the tunnel entrance area size parameter, which is the set value; is the tunnel exit area size parameter, which is the set value.

[0033] Furthermore, compensation strength Calculated by the following formula:

[0034] ;

[0035] in, is the initial compensation strength, ranging from 0.4 to 0.6; is the maximum brightness of the intermediate image; is the minimum brightness of the intermediate image; is the brightness scale factor, ranging from 0.05 to 0.15; is the standard deviation of the pixel values ​​of the intermediate image; is the reference standard deviation of the pixel values ​​of the intermediate image, and is the set value; is the entropy of the intermediate image; is the entropy adjustment coefficient, which is an integer from 1 to 3; is the reference standard entropy, which is an integer from 6 to 10; is the distance from the pixel of the current intermediate image to the nearest tunnel entrance or exit; is half the length of the middle image diagonal.

[0036] Furthermore, in step 3, the adaptive threshold is applied to the compensated image by the following formula to obtain the denoised image :

[0037] ;

[0038] in, is the pixel position set of the compensated image; is the scale number of the compensated image.

[0039] By adopting the above technical scheme, the present invention produces the following beneficial effects: the present invention significantly improves the denoising effect of the image by introducing adaptive threshold calculation and multi-scale denoising method. In low-light scenes, the noise level is high and complex. Through the adaptive threshold calculation method, the thresholds of different scales and positions can be dynamically adjusted to suppress the noise more accurately. This adaptability ensures that noise can be effectively removed under different lighting conditions to improve the clarity and quality of the image. In the denoising process, detail retention is a key challenge. Traditional denoising methods often blur image details while removing noise, resulting in a decrease in image quality. The present invention effectively balances the contradiction between denoising and detail retention by introducing multi-scale analysis and adaptive thresholds. The multi-scale analysis method can process images at different resolutions, so that high-frequency details are retained and low-frequency noise is effectively removed. Through the dynamic adjustment of the adaptive threshold, the fine processing of the detail-rich area is ensured, so that the image still maintains high resolution and detail clarity after denoising. In low-light scenes such as night and tunnels, the lighting conditions are complex and changeable, and the traditional fixed threshold method is difficult to adapt to such changes. The present invention dynamically calculates the adaptive threshold by comprehensively considering factors such as brightness contrast, noise level, image entropy and spatial position, making the noise reduction process highly adaptive. Whether it is a night scene with extremely weak light or a tunnel entrance and exit with rapidly changing light, the method of the present invention can automatically adjust the parameters to ensure the best noise reduction effect and image quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1A schematic diagram of a method flow of a deep learning denoising method for images in low-light scenes at night and in tunnels provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0041] All features disclosed in this specification, or steps in all methods or processes disclosed, except mutually exclusive features and / or steps, can be combined in any manner.

[0042] Any feature disclosed in this specification (including any additional claims and abstract), unless otherwise stated, may be replaced by other alternative features that are equivalent or have similar purposes. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0043] Example 1: Reference Figure 1 , a deep learning denoising method for images in low-light scenes at night and in tunnels, the method comprising:

[0044] Step 1: After nonlinear mapping of the target low-light scene image, adaptive enhancement is performed to obtain an intermediate image;

[0045] The present invention performs nonlinear mapping on the target low-light scene image. Nonlinear mapping is an important preprocessing step, which adjusts the brightness and contrast of the image by applying a nonlinear function to better adapt to the image characteristics under low-light conditions. Specifically, the nonlinear mapping function can be a logarithmic transformation, a gamma correction, etc. These nonlinear functions can enhance the details in the image while maintaining the overall structure of the image, making the low-brightness part clearer. For example, the logarithmic transformation can expand the dynamic range of the low-brightness area by taking the logarithm of the image grayscale value, while the gamma correction can change the brightness curve of the image by adjusting the gamma value, thereby enhancing the image contrast without increasing noise. After completing the nonlinear mapping, the present invention uses an adaptive enhancement technology to further process the mapped image. Adaptive enhancement is a method that can adaptively adjust the enhancement parameters according to the local features of the image. Commonly used technologies include histogram equalization and contrast-limited adaptive histogram equalization (CLAHE). The core idea of ​​adaptive enhancement is to dynamically adjust the brightness and contrast of the image by analyzing the local statistical features of the image, so that the enhancement effect is more natural and smooth. Histogram equalization redistributes the grayscale values ​​of an image to make the grayscale histogram of the image more uniform, thereby improving the contrast of the image. CLAHE introduces the concept of local enhancement based on histogram equalization, and avoids noise amplification and artifacts caused by excessive enhancement by limiting the enhancement amplitude of local contrast.

[0046] The specific process of adaptive enhancement is as follows: First, the image is divided into several small blocks, and each small block is subjected to histogram equalization processing. Then, the processed small blocks are reassembled into a complete image. In this process, the CLAHE algorithm controls the distribution of pixel values ​​in each small block by setting a contrast limit threshold, so that the enhancement effect is more uniform and smooth. In addition, the CLAHE algorithm also smoothes the boundaries between small blocks by interpolation method to avoid the generation of image block effect. Through the dual processing of nonlinear mapping and adaptive enhancement, the present invention can effectively improve the brightness and contrast of images in low-light scenes, enhance image details, and provide higher quality input images for subsequent wavelet transform, multi-scale decomposition and noise reduction processing. This processing step not only improves the visual effect of the image, but also provides more reliable and clear features for the subsequent deep learning model, thereby improving the performance and effect of the entire image denoising system. Compared with traditional image enhancement methods, the nonlinear mapping and adaptive enhancement methods proposed in the present invention can better adapt to the special requirements of low-light scenes, significantly improve the quality of the image and the ability to retain details, and lay a solid foundation for achieving efficient image denoising.

[0047] Step 2: Use wavelet transform to decompose the intermediate image into multiple scales and estimate the noise standard deviation at each scale;

[0048] Wavelet transform is a powerful time-frequency analysis tool that can decompose an image into sub-bands of different scales, each of which contains information about the image in different frequency ranges. Compared with the traditional Fourier transform, wavelet transform has localized characteristics and can provide information in both the time domain and the frequency domain, which makes it particularly suitable for processing local details and edge information contained in the image. In low-light scenes, images usually contain a large amount of high-frequency noise, which is mainly concentrated in the high-frequency sub-bands of wavelet transform. Therefore, through wavelet transform, the noise component in the image can be effectively separated from the useful signal component. Specifically, the intermediate image is first subjected to wavelet transform to decompose it into sub-bands of different scales. Each sub-band of each scale contains the detail information and noise information of the image at that scale. Through this multi-scale decomposition, the characteristics of the image can be observed at different resolutions, so as to better analyze and process the noise. After completing the wavelet transform, the noise standard deviation needs to be estimated at each scale. The noise standard deviation is an important parameter to measure the noise intensity, which directly affects the subsequent noise reduction processing effect.

[0049] In order to estimate the noise standard deviation, the high-frequency subband of the image can be used. Generally speaking, the high-frequency subband mainly contains the detailed information and noise of the image, while the low-frequency subband mainly contains the outline and overall structure of the image. In the high-frequency subband, the noise component is usually manifested as random high-frequency vibration, so the noise standard deviation can be effectively estimated by statistical analysis of the high-frequency subband. Commonly used methods include the median absolute deviation (MAD) method and other statistical methods, which can accurately estimate the noise standard deviation without relying on a specific noise model. Through multi-scale decomposition and noise standard deviation estimation, the noise characteristics of the image can be obtained at different scales and frequency ranges. This provides an important basis for subsequent noise reduction processing. In low-light scenes, due to the poor lighting conditions of the image, the noise intensity is often large, so accurately estimating the noise standard deviation is crucial to achieve effective noise reduction. The present invention can extract the noise characteristics of the image at different resolutions through wavelet transform and multi-scale decomposition, so as to apply adaptive noise reduction strategies at different scales to achieve better noise reduction effects. In addition, the present invention also takes into account the influence of ambient light intensity on noise. Under different lighting conditions, the intensity and characteristics of noise will be different. Therefore, when estimating the noise standard deviation, the information of ambient light intensity is combined to more accurately reflect the actual noise situation. This noise estimation method combined with ambient light intensity enables the noise reduction algorithm to better adapt to different lighting conditions, thereby achieving good noise reduction effects in various low-light scenes.

[0050] Step 3: Calculate the adaptive threshold at each scale and spatial position based on the estimated noise standard deviation and ambient light intensity; Consider the tunnel effect, compensate the tunnel entrance area and tunnel exit area corresponding to the sharp change in brightness in the intermediate image to obtain a compensated image, and apply the adaptive threshold at each scale spatial position to the compensated image to obtain a denoised image.

[0051] The adaptive threshold for each scale and spatial position can be calculated by the noise standard deviation and ambient light intensity obtained in the previous step. The calculation of the adaptive threshold is based on the comprehensive consideration of noise characteristics and lighting conditions, so that the appropriate threshold can be used for noise removal in different image areas and scales. This ensures that a higher threshold is used in high-noise areas to effectively remove noise, while a lower threshold is used in low-noise areas to retain more image details. Specifically, the calculation formula of the adaptive threshold combines the noise standard deviation and light intensity to dynamically adjust the threshold of each position, so that the threshold can flexibly adapt to different parts of the image to achieve the best noise reduction effect. After calculating the adaptive threshold, the impact of the tunnel effect on the image must also be considered. The tunnel entrance and exit areas usually experience a sharp change in brightness, which makes it difficult for traditional noise reduction methods to distinguish between noise and useful signals, thereby affecting the noise reduction effect. In order to solve this problem, the present invention introduces a brightness compensation mechanism in the tunnel entrance and exit areas. By analyzing the brightness changes of the image, brightness compensation is performed on the tunnel entrance and exit areas to ensure that the image details in these areas can be correctly identified and retained. The specific implementation of brightness compensation can be achieved by adjusting the pixel values ​​of these areas to make the brightness change smoother, thereby reducing the possibility of misjudging noise.

[0052] After completing the brightness compensation, the adaptive threshold is applied to the compensated image for noise reduction. The core idea of ​​noise reduction is to compare the value of each pixel with the adaptive threshold to determine whether the pixel is noise and perform corresponding processing. If the pixel value exceeds the adaptive threshold, the pixel value is retained; otherwise, it is regarded as noise and smoothed or other noise reduction operations are performed. This noise reduction method based on the adaptive threshold can effectively remove noise while retaining the image details, thereby improving the quality and clarity of the image. In this way, the present invention can achieve efficient image noise reduction under different lighting conditions and scenes. In low-light scenes such as night and tunnels, images are usually subject to large noise interference, and through the calculation and application of the adaptive threshold, the noise reduction strategy can be flexibly adjusted to ensure that the best effect can be achieved in different regions and scales. In addition, the brightness compensation mechanism of the tunnel effect enables the effective processing of areas with sharp brightness changes, further improving the noise reduction effect and image quality. This step of the present invention retains the details and structural information of the image while achieving accurate noise reduction, significantly improving the overall visual effect of the image. Compared with traditional noise reduction methods, the present invention can better adapt to the special requirements of low-light scenes by introducing adaptive thresholds and brightness compensation mechanisms, providing more natural and clear images. Through this innovative noise reduction method, the present invention not only improves the noise reduction effect of the image, but also enhances the detail retention ability of the image, laying a solid foundation for achieving high-quality image processing. The application of this technology will help improve the image processing quality in low-light scenes, and is widely used in night monitoring, automatic driving, tunnel detection and other fields, with important practical application value and broad market prospects.

[0053] Embodiment 2: The low-light scene image is an image acquired in an environment where the light intensity is less than or equal to 100 lux.

[0054] Specifically, images acquired in an environment with a light intensity of less than or equal to 100 lux usually have obvious noise interference. This is because under low-light conditions, the camera sensor needs a longer exposure time or an increased sensitivity (ISO) to capture enough light. However, both methods will lead to an increase in noise. Long exposure time is prone to introduce motion blur and thermal noise, while increasing ISO will amplify the electronic noise of the sensor. Therefore, the noise problem of images acquired under these low-light conditions is particularly serious, which poses a great challenge to subsequent image processing.

[0055] Embodiment 3: The target low-light scene image is analyzed by the following formula: Perform nonlinear mapping to obtain a nonlinear mapping image :

[0056] ;

[0057] In the nonlinear mapping formula The purpose of this part is to expand the dynamic range of an image. In low light conditions, the brightness values ​​of an image are usually low, making it difficult to discern details. The logarithmic transformation can effectively handle this situation by taking the logarithm to stretch the lower brightness values, making the dark details more obvious. Specifically, when When smaller, It will significantly increase the brightness of the image, so that the details of the low-brightness area can be enhanced without introducing too much noise. This logarithmic transformation feature makes it very suitable for low-light scenes and can significantly improve the visibility of the image. Next, the formula is a typical S-shaped function, which is used to smoothly transition between the low-brightness and high-brightness parts of the image, thereby avoiding abrupt brightness changes. This part achieves nonlinear adjustment of image brightness through the combination of exponential function and denominator. Specifically, when When smaller, Close to , and when When is large, the value of this part approaches 1, which effectively compresses the dynamic range of the high-brightness part and makes the overall brightness distribution of the image more uniform and smooth. This smooth transition effect is very important for processing brightness changes in low-light scenes, and can effectively reduce brightness mutations and edge effects in the image. Next, in the formula Combining the hyperbolic tangent function and the sine function further enhances the local contrast and details of the image. It is an activation function commonly used in neural networks, which can enhance the local features of the image without introducing too much noise. This increases the periodic changes in the image, making the image details richer and more vivid. This combination enhances the layering and details of the image by superimposing features of different frequencies, making the image in low-light scenes visually clearer and more textured. In addition, The power function is a combination of the power function and the error function. In low light conditions, it can effectively increase the brightness of the image, making the originally dim area brighter. At the same time, the error function It has smooth characteristics and can effectively suppress noise through its characteristics in low-light areas. The introduction of the error function makes the image enhancement effect more natural and avoids the noise amplification problem introduced by excessive brightness enhancement. This combination allows the image to retain good smoothness and naturalness while enhancing the brightness, improving the visual quality of the image. The last part, It combines the exponential decay function and the fractional form. This is done by the exponential decay function The low-brightness part of the image is enhanced, making the previously invisible details visible. According to the global average brightness Dynamically adjust the degree of image enhancement to ensure that a reasonable enhancement effect can be obtained under different brightness conditions. The characteristics of the exponential decay function make the image enhancement effect nonlinear, which can effectively improve the brightness under low light conditions, and the introduction of the fractional form ensures the smoothness and naturalness of the overall enhancement effect.

[0058] in, is the target low-light scene image, with a value range of [0,1]; for The global average brightness of is the error function, and its specific form is as follows:

[0059] ;

[0060] In low light conditions, images contain a lot of noise, especially in darker areas, where the impact of noise is particularly significant. Traditional image enhancement methods tend to introduce more noise when processing these areas, causing image quality to deteriorate. The error function The introduction of provides an effective solution to this problem. By applying the error function to the pixel values ​​of the image, its smoothing characteristics can be used to effectively suppress noise while enhancing the details and contrast of the image. Specifically, the smoothing characteristic of the error function is that it gradually rises in the low value range and tends to be stable in the high value range. This characteristic allows it to enhance the low brightness area while not excessively amplifying the noise in the high brightness area. When smaller, The output value of is relatively small, and the image enhancement effect is relatively mild, which helps to suppress random noise in low light conditions. When larger, The output value of tends to 1. This smooth transition effectively avoids the problem of over-enhancement and noise amplification in high-brightness areas.

[0061] in, is the independent variable of the error function; is the integration variable.

[0062] Embodiment 4: In step 1, nonlinear mapping image Perform adaptive enhancement to obtain the intermediate image The formula is as follows:

[0063] ;

[0064] in, yes The global pixel mean of ; yes The global pixel standard deviation of for The global average brightness of and They are The width and height of and The horizontal and vertical coordinates of the pixels are ； is an adaptive parameter, and its calculation formula is as follows:

[0065] ;

[0066] is the gamma correction parameter, calculated as follows:

[0067] .

[0068] Specifically, the formula is an adaptive parameter, and its calculation formula is ,here is the global average brightness of the nonlinear mapping image. When it is close to 0.5, Take the middle value to make the enhancement effect smoothly transition between the two extreme states; When the deviation is 0.5, Then it tends to 0 or 1, thereby dynamically adjusting the enhancement weights so that the enhancement effect is more adaptive to the global brightness characteristics of the image. Next, the first part of the formula The Gaussian function is used to enhance the image. Here, yes The global pixel mean, is the global pixel standard deviation. The introduction of the Gaussian function adjusts the Gaussian distribution of pixel values ​​so that the pixels near the mean are enhanced. This processing method can highlight the details of the image while maintaining the overall brightness distribution of the image, especially in low light conditions, by enhancing the pixels in the medium brightness area, improving the visual effect of the image and effectively suppressing the influence of noise.

[0069] The second part of the formula The sine function is used to periodically enhance the image in the spatial domain. Here, and are the width and height of the image, respectively. and is the horizontal and vertical coordinates of the pixel. The introduction of the sine function enhances the local contrast and details of the image by periodically modulating the image in the spatial domain. This processing method can improve the texture details and visual clarity of the image without significantly increasing the noise. It is particularly suitable for processing images in low-light environments, making the image details richer and more layered. Finally, the third part of the formula It combines the idea of ​​gamma correction. Here, is the global average brightness of the nonlinearly mapped image, is the gamma correction parameter. Through this nonlinear transformation, the low-brightness area of ​​the image can be more strongly enhanced, while the high-brightness area can be less enhanced, thus avoiding the overexposure problem of the high-brightness area. Gamma correction is widely used in image processing and can effectively improve the overall brightness and contrast of the image, making the image more natural and realistic visually.

[0070] The core idea of ​​the adaptive enhancement method is to achieve dynamic adjustment of different regions and features of the image through complex nonlinear formulas. Images in low-light scenes usually have large brightness differences and significant noise problems, so a flexible processing method is needed to balance the enhancement effect and noise suppression. and gamma correction parameters , so that the enhancement effect can be dynamically adjusted according to the global and local features of the image, ensuring that good processing effects can be achieved under different lighting conditions. In addition, the adaptive enhancement method not only improves the details and contrast of the image, but also ensures that good processing effects can be achieved under different lighting conditions through a combination of complex mathematical functions. Compared with traditional image enhancement methods, the present invention can better adapt to the special needs of low-light scenes by combining multiple nonlinear processing technologies, significantly improve the quality and visual effects of images, and provide high-quality input data for subsequent deep learning noise reduction algorithms. This method is not only innovative in theory, but also shows strong adaptability and superior image enhancement effects in practical applications, with important practical application value and broad market prospects. Through this innovative processing method, the present invention realizes efficient image enhancement and noise reduction in low-light scenes, providing important technical support and guarantee for practical applications. In extreme low-light environments such as nighttime and tunnels, images usually have problems such as low brightness, high noise, and loss of details due to insufficient light. By introducing the adaptive enhancement method, the enhancement parameters and processing methods can be dynamically adjusted according to the global brightness characteristics and local detail characteristics of the image, making the enhancement effect more flexible and efficient.

[0071] Embodiment 5: After performing multi-scale decomposition on the intermediate image using wavelet transform in step 2, the noise standard deviation is estimated at each scale using the following formula:

[0072] ;

[0073] in, For the Scale at light level The standard deviation of noise under For the The scale of wavelet coefficients; For the The mean of the wavelet coefficients at each scale; For the The number of wavelet coefficients at each scale; is the set adjustment parameter; is the central scale; For the The maximum value of the wavelet coefficients at each scale.

[0074] Specifically, the preliminary noise standard deviation is obtained by calculating the sum of squares of the difference between the wavelet coefficients and their mean at each scale. This part reflects the advantages of traditional statistical methods, which can independently analyze noise characteristics at different scales. In the formula, Representative The first scale wavelet coefficients, is the mean of the wavelet coefficients at this scale. This method preliminarily quantifies the noise level at this scale through statistical analysis of each wavelet coefficient. However, it is difficult to fully reflect the complex changes of noise under low light conditions by relying solely on traditional statistical methods. Therefore, the formula further introduces the light level The relevant logarithmic adjustment term. The specific form is ,here Representative The introduction of this logarithmic term, by taking the logarithm of the ratio of the absolute value of the wavelet coefficient to its maximum value, can dynamically adjust the estimate of the noise standard deviation, so that it can better adapt to the noise characteristics under different lighting conditions. In low-light conditions, noise is often more significant, and the logarithmic term can effectively enhance the perception and estimation of these noises, thereby improving the accuracy of noise processing. In addition, in order to consider the correlation between scales and the impact of lighting changes on noise, the Gaussian attenuation function is also introduced in the formula .in, is the central scale, is the adjustment parameter. The role of this Gaussian decay function is that as the scale Away from center scale , its influence gradually weakens, thereby ensuring the smoothness and continuity of the noise standard deviation estimation. In this way, the details and features of the image at different scales can be better captured, while effectively reducing the impact of noise. Through this multi-scale, multi-parameter comprehensive analysis method, the present invention can dynamically adjust the noise estimation parameters under extreme low light conditions, thereby performing fine noise suppression processing at all scales. Compared with traditional single-scale noise processing methods, this multi-scale method can more comprehensively reflect the multi-level characteristics and noise distribution of the image, making the noise reduction effect more natural and efficient.

[0075] Embodiment 6: Calculate the adaptive threshold for each scale and spatial position using the following formula:

[0076] ;

[0077] in, For the At this scale, pixel position Adaptive threshold of is the reference light intensity, equal to 100 lux; pixel position , is the horizontal axis, is the vertical coordinate; the center pixel position , is the horizontal coordinate of the center pixel position; is the ordinate of the center pixel position; , , and are control parameters, among which, is an integer ranging from 2 to 4; The value range is from 0 to 1; The value range of is 0.1 to 0.3; ; To find the determinant of a matrix.

[0078] Specifically, the first term in the formula is ,in is a control parameter whose value range is an integer from 2 to 4 and is used to adjust the overall amplitude of the threshold. is a constant term related to the image size, represents the total number of pixels in the image. The purpose of this term is to ensure that the threshold calculation is adapted to the image size, so that a reasonable threshold can be obtained in images of different sizes. Next, the formula The effect of light intensity is partially introduced. Here, is the reference light intensity, equal to 100 lux. The lower it is, the greater the impact of this term, meaning that in low-light conditions, the threshold will be adjusted according to the light level, thereby enhancing the noise reduction effect. is a control parameter whose value range is from 0 to 1. The value of can control the degree of influence of light intensity on the threshold. Noise is usually more significant under low light conditions, so through this adjustment, the noise reduction effect can be enhanced in low light conditions to ensure improved image quality. Finally, the formula The effect of spatial position is partially considered. Pixel position Expressed as ,in is the horizontal axis, is the vertical axis, is the center pixel position. The introduction of this part enables the adaptive threshold to be adjusted according to the spatial position. is a control parameter, whose value range is 0.1 to 0.3. The value of can control the influence of spatial position on the threshold. is a parameter related to the image size and is defined as ,in and are the width and height of the image respectively. Through this spatial position adjustment, different thresholds can be applied to different areas of the image, especially in the case of uneven noise distribution, this adjustment is particularly important. The various items in the formula work together to enable the calculation of the adaptive threshold to flexibly adapt to different environments and conditions. In low-light scenes such as night and tunnels, images usually generate a lot of noise due to insufficient light, and the noise distribution is complex and changeable. Through the above formula, the threshold can be dynamically adjusted at each scale and spatial position to ensure that the best noise reduction effect can be achieved under different conditions. The adaptive threshold calculation method of the present invention provides a flexible and efficient image denoising solution by combining multiple factors such as noise level, light intensity and spatial position. Compared with the traditional fixed threshold method, this adaptive method can better handle complex noise distribution, especially under extreme low light conditions, by dynamically adjusting the threshold, the quality and clarity of the image can be significantly improved. The introduction of the adaptive threshold not only improves the detail retention ability of the image, but also effectively reduces the impact of noise on the image, making the processed image more natural and realistic.

[0079] Embodiment 7: In step 3, taking into account the tunnel effect, the tunnel entrance area and tunnel exit area corresponding to the sharp change in brightness in the intermediate image are compensated by the following formula to obtain a compensated image :

[0080] ;

[0081] in, is the preset tunnel effect compensation function, which is calculated using the following formula:

[0082] ;

[0083] in, To compensate for the strength; are the coordinates of the center of the tunnel entrance, is the horizontal axis, is the vertical coordinate; is the center coordinate of the tunnel exit, is the horizontal axis, is the vertical coordinate; is the tunnel entrance area size parameter, which is the set value; is the tunnel exit area size parameter, which is the set value;

[0084] Specifically, the impact of the tunnel effect is that at the entrance of the tunnel, the image brightness will drop sharply due to the sudden entry from a bright environment into a darker tunnel, resulting in blurred details and obvious noise. Similarly, at the exit of the tunnel, the image brightness will rise sharply when entering a bright environment from a dark tunnel, causing similar problems. In order to solve these problems, the present invention proposes a compensation formula to obtain a compensated image by processing the intermediate image. The core of the compensation formula lies in the compensation function , this function is calculated based on the geometric position and brightness distribution characteristics of the tunnel entrance and exit. Specifically, the compensation function simulates the brightness changes in the tunnel entrance and exit areas in the form of a Gaussian function. The Gaussian function has a smooth transition characteristic and can effectively simulate the gradual change of brightness in the actual environment, thereby avoiding the adverse effects caused by the sharp change of brightness in the image.

[0085] The Gaussian function is used to achieve a smooth transition of brightness in space. For example, the Gaussian function in the tunnel entrance area ,in are the center coordinates of the tunnel entrance, is the size parameter of the tunnel entrance area. This function reaches its maximum value at the center of the entrance and gradually decreases as the distance from the center point increases, simulating the brightness change when entering the tunnel from a bright environment. Similarly, the Gaussian function of the tunnel exit area ,in are the center coordinates of the tunnel exit, is the size parameter of the tunnel exit area. This function reaches its maximum value at the exit center and gradually decreases as the distance from the center point increases, simulating the brightness change from a dark tunnel into a bright environment. Compensation function By taking the maximum value of the two, comprehensive compensation for the tunnel entrance and exit areas is achieved. This design ensures that the brightness transition in the tunnel entrance and exit areas can be effectively processed, thereby reducing the loss of details and noise caused by the sharp change in brightness. is the compensation strength parameter, which is used to control the strength of compensation. The value of can be adjusted flexibly according to the specific image conditions to achieve the best image quality. The value of can achieve the best brightness compensation effect while ensuring image details and quality. The compensation formula is designed to achieve dynamic brightness adjustment to ensure smooth brightness transition of the image in the tunnel entrance and exit areas, thereby improving the overall image quality. By compensating for these key areas, the adverse effects of brightness changes can be significantly reduced, and the details and clarity of the image can be improved. The advantage of this tunnel effect compensation method lies in its flexibility and adaptability. By adjusting the parameters of the compensation function, it can adapt to different tunnel structures and lighting conditions, so that good processing effects can be achieved in various complex environments. For example, in tunnels of different lengths and widths, by adjusting and The value of can accurately control the size of the tunnel entrance and exit areas, thereby achieving targeted compensation. The value of can control the compensation intensity, making image processing more flexible.

[0086] The tunnel effect compensation method of the present invention significantly improves the image quality in low-light scenes such as nighttime and tunnels by modeling and compensating for the brightness changes in the tunnel entrance and exit areas. In practical applications, this method can effectively improve the visual effect of the image, making the image clearer and more natural. Compared with traditional image processing methods, the method of the present invention not only considers the adjustment of the overall brightness, but also carefully processes the local brightness changes, thereby achieving a more uniform and natural image effect.

[0087] Example 8: Compensation Strength Calculated by the following formula:

[0088] ;

[0089] in, is the initial compensation strength, ranging from 0.4 to 0.6; is the maximum brightness of the intermediate image; is the minimum brightness of the intermediate image; is the brightness scale factor, ranging from 0.05 to 0.15; is the standard deviation of the pixel values ​​of the intermediate image; is the reference standard deviation of the pixel values ​​of the intermediate image, and is the set value; is the entropy of the intermediate image; is the entropy adjustment coefficient, which is an integer from 1 to 3; is the reference standard entropy, which is an integer from 6 to 10; is the distance from the pixel of the current intermediate image to the nearest tunnel entrance or exit; is half the length of the middle image diagonal.

[0090] Specifically, the initial compensation intensity in the formula is a basic parameter with a value range of 0.4 to 0.6, which provides a basic compensation strength. This initial value ensures that the compensation process can start working without further adjustment. By dynamically adjusting other parameters, the formula can flexibly change the compensation strength under different conditions to achieve the best effect. The first part of the formula reflects the brightness contrast of the image. Here, and are the maximum and minimum brightness of the middle image, is the brightness scale factor, ranging from 0.05 to 0.15. Brightness contrast is an important indicator to measure the brightness difference between different areas of an image, which directly affects the visual effect and detail performance of the image. By calculating the brightness contrast and using the hyperbolic tangent function After smoothing, the formula can normalize the brightness contrast to between 0 and 1. This process ensures that the compensation intensity can be reasonably adjusted even when the brightness changes drastically, thus avoiding over-compensation or under-compensation. The brightness standard deviation of the image is introduced and reference standard deviation . The brightness standard deviation of an image is an important indicator to measure the uniformity of the brightness distribution of the image. A larger standard deviation means that there is a larger brightness variation in the image, which usually causes more noise and loss of details. By calculating the ratio of the standard deviations and processing it with an exponential decay function, the formula can effectively adjust the compensation intensity to adapt it to the actual brightness distribution of the image. If the standard deviation is large, the compensation intensity will increase to better deal with the problems caused by brightness variations; conversely, if the standard deviation is small, the compensation intensity will decrease to avoid over-compensation. Part III Combined with the entropy of the image and the reference standard entropy . Entropy is an important indicator to measure the amount of information and complexity of an image, reflecting the distribution of different gray levels in the image. A higher entropy value means that the image contains more details and information, while a lower entropy value means that the image is simpler. By calculating the image entropy and smoothing it with the logistic function, the formula can dynamically adjust the compensation strength according to the complexity of the image. Parameters is the entropy adjustment coefficient, which is an integer from 1 to 3 and is used to control the effect of entropy on the compensation strength. If the image entropy value is high, the formula will increase the compensation strength to ensure that the image details are fully preserved; conversely, if the entropy value is low, the compensation strength will be reduced to avoid unnecessary compensation. The last part The distance from the pixel in the image to the nearest tunnel entrance or exit is considered The impact on the compensation strength. Here, It is half the length of the diagonal of the intermediate image. The distance function is introduced to deal with areas with sharp changes in brightness in the tunnel effect. By using the square form of the sine function, the formula can increase the compensation intensity when the image approaches the entrance or exit area of ​​the tunnel, thereby effectively smoothing the brightness transition in these areas. In this way, areas with large brightness changes in the image can receive stronger compensation, while areas with small brightness changes have their compensation intensity reduced accordingly to maintain the overall balance and naturalness of the image.

[0091] By combining multiple image features and environmental factors, the compensation intensity is dynamically adjusted, thereby achieving efficient processing of images in low-light scenes. By comprehensively considering brightness contrast, brightness standard deviation, image entropy and pixel distance, the formula ensures the rationality and adaptability of the compensation intensity. In practical applications, this method can effectively improve the quality of images in low-light environments such as night and tunnels, making the images clearer and more natural. The method of the present invention demonstrates its strong flexibility and adaptability when processing low-light images. By dynamically adjusting the compensation intensity, it can provide the best image processing effect under different lighting conditions and image characteristics. This innovative processing method not only improves the visual effect of the image, but also provides high-quality input data for the deep learning denoising algorithm, thereby further improving the performance and effect of the overall image processing system. In addition, this method has a wide range of applicability in practical applications. By adjusting the parameters in the formula, it can adapt to different tunnel structures and lighting conditions, so that good processing effects can be achieved in various complex environments. For example, in a long tunnel or a short tunnel, by adjusting the distance parameter and brightness scale factor , the compensation intensity can be precisely controlled, making image processing more flexible and efficient. Similarly, by adjusting the initial compensation intensity and entropy adjustment coefficient , the influence of different image features on the compensation intensity can be controlled to meet the needs of different application scenarios.

[0092] Example 9: In step 3, the adaptive threshold is applied to the compensated image to obtain the denoised image by the following formula: :

[0093] ;

[0094] in, is the pixel position set of the compensated image; is the scale number of the compensated image.

[0095] Specifically, in the formula, represents the final denoised image, is the image enhanced by the above steps, and is the image after tunnel effect compensation. Applied to these images, accurate noise suppression can be achieved. Specifically, represents the set of pixel positions of the compensated image, is the scale number of the compensated image. Each term in the formula reflects the process of applying adaptive thresholds to adjust the image at different scales and positions. It is calculated based on multiple features of the image in the previous steps. These features include brightness contrast, noise level, local details of the image, and spatial position. The introduction of adaptive thresholds enables the denoising process to dynamically adapt to different areas and brightness conditions of the image, providing appropriate noise suppression strength for each pixel position. In this way, noise can be effectively removed while avoiding the problem of detail loss caused by over-smoothing. The symbolic function in the formula It plays a role in determining the direction of adjustment. This sign function determines whether a certain pixel needs to be adjusted based on the difference between the compensated image and the enhanced image. The output of the sign function is +1 or -1, indicating whether the threshold of the pixel needs to be increased or decreased. In this way, the formula can dynamically adjust the value of each pixel to achieve more accurate noise suppression. In low-light scenes, images are usually subject to greater noise interference, especially in the tunnel entrance and exit areas, where drastic changes in lighting conditions make it easy for obvious noise and detail loss to appear in the image. By applying the adaptive threshold method, smoother brightness transitions and more accurate noise suppression can be achieved in these key areas. Specifically, the calculation of the adaptive threshold takes into account a variety of factors, including the local brightness contrast of the image, the global brightness distribution, the noise level, and the spatial characteristics of the image. These factors work together to enable the application of the threshold to be accurate to every pixel position and every scale, thereby achieving the best noise reduction effect. In practical applications, the parameters and calculation steps in the formula can be adjusted according to the specific image characteristics and processing requirements. For example, for images with high noise levels, the weight of the adaptive threshold can be appropriately increased to more effectively suppress noise. For images with more details, the threshold can be applied less strongly to retain more details and texture information.

[0096] Although the specific embodiments of the present invention are described above, it should be understood by those skilled in the art that these specific embodiments are only illustrative, and those skilled in the art may omit, replace, and change the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, merging the above method steps so as to perform substantially the same functions in substantially the same manner to achieve substantially the same results is within the scope of the present invention. Therefore, the scope of the present invention is limited only by the appended claims.

Claims

1. A deep learning denoising method for images in low-light scenes at night and in tunnels, characterized in that: The method comprises: Step 1: After nonlinear mapping of the target low-light scene image, adaptive enhancement is performed to obtain an intermediate image; Step 2: Use wavelet transform to decompose the intermediate image into multiple scales and estimate the noise standard deviation at each scale; Step 3: Calculate the adaptive threshold at each scale and spatial position based on the estimated noise standard deviation and ambient light intensity; Consider the tunnel effect, compensate the tunnel entrance area and tunnel exit area corresponding to the sharp change in brightness in the intermediate image to obtain a compensated image, and apply the adaptive threshold at each scale spatial position to the compensated image to obtain a denoised image; In step 3, the tunnel effect is taken into account, and the tunnel entrance area and tunnel exit area corresponding to the sharp change in brightness in the intermediate image are compensated by the following formula to obtain the compensated image : ； in, is the middle image; and are the horizontal and vertical coordinates of the pixel respectively; is the preset tunnel effect compensation function, which is calculated using the following formula: ； in, To compensate for the strength; are the coordinates of the tunnel entrance center, is the horizontal axis, is the vertical coordinate; is the center coordinate of the tunnel exit, is the horizontal axis, is the vertical coordinate; is the tunnel entrance area size parameter, which is the set value; is the tunnel exit area size parameter, which is the set value; Compensation Strength Calculated by the following formula: ； in, is the initial compensation strength, ranging from 0.4 to 0.6; is the maximum brightness of the intermediate image; is the minimum brightness of the intermediate image; is the brightness scale factor, ranging from 0.05 to 0.15; is the standard deviation of the pixel values ​​of the intermediate image; is the reference standard deviation of the pixel values ​​of the intermediate image, and is the set value; is the entropy of the intermediate image; is the entropy adjustment coefficient, which is an integer from 1 to 3; is the reference standard entropy, which is an integer from 6 to 10; is the distance from the pixel of the current intermediate image to the nearest tunnel entrance or exit; is half the length of the middle image diagonal.

2. The deep learning denoising method for images in low-light scenes at night and in tunnels according to claim 1, characterized in that: The low-light scene image is an image acquired in an environment where the light intensity is less than or equal to 100 lux.

3. The deep learning denoising method for images in low-light scenes at night and in tunnels as claimed in claim 2, characterized in that: The target low-light scene image is obtained by the following formula: Perform nonlinear mapping to obtain a nonlinear mapping image : ； in, is the target low-light scene image, with a value range of [0,1]; for The global average brightness of is the error function, and its specific form is as follows: ； in, is the independent variable of the error function; is the integration variable.

4. The deep learning denoising method for images in low-light scenes at night and in tunnels as claimed in claim 3, characterized in that: In step 1, the nonlinear mapping image Perform adaptive enhancement to obtain the intermediate image The formula is as follows: ； in, yes The global pixel mean of ; yes The global pixel standard deviation of for The global average brightness of and They are The width and height of is an adaptive parameter, and its calculation formula is as follows: ； is the gamma correction parameter, calculated as follows: 。 5. The deep learning denoising method for images in low-light scenes at night and in tunnels as claimed in claim 4, characterized in that: Step 2: After multi-scale decomposition of the intermediate image using wavelet transform, the noise standard deviation is estimated at each scale using the following formula: ； in, For the Scale at light level The standard deviation of noise under For the The scale of wavelet coefficients; For the The mean of the wavelet coefficients at each scale; For the The number of wavelet coefficients at each scale; is the set adjustment parameter; is the central scale; For the The maximum value of the wavelet coefficients at each scale.

6. The deep learning denoising method for images in low-light scenes at night and in tunnels as claimed in claim 5, characterized in that: The adaptive threshold is calculated for each scale and spatial position using the following formula: ； in, For the At this scale, pixel position Adaptive threshold of is the reference light intensity, equal to 100 lux; pixel position , is the horizontal axis, is the vertical coordinate; the center pixel position , is the horizontal coordinate of the center pixel position; is the ordinate of the center pixel position; , , and are control parameters, among which, is an integer ranging from 2 to 4; The value range is from 0 to 1; The value range of is 0.1 to 0.3; ; To find the determinant of a matrix; Represents the total number of pixels in the image.

7. The deep learning denoising method for images in low-light scenes at night and in tunnels according to claim 6, characterized in that: In step 3, the adaptive threshold is applied to the compensated image to obtain the denoised image using the following formula: : ; in, is the pixel position set of the compensated image; is the scale number of the compensated image.

Images