Dark light image enhancement method and system fusing deep learning, and medium

Abstract

The application relates to the technical field of digital image processing, in particular to a dark-light image enhancement method and system fusing deep learning and a medium, and specifically comprises the following steps: collecting a building material surface RGB image under a dark-light environment; obtaining each type of pixel point based on dark channel value gradient division; analyzing the change difference degree of the response of the pixel points in each connected domain to the light intensity on the RGB three channels under each type of pixel point to adjust the estimated intensity of the light component obtained by Gaussian filtering in the Retinex algorithm; analyzing the relative size of the size of each connected domain in the connected domain of the same light type and the local texture change feature to adjust the Gaussian filtering kernel size, completing the construction of an adaptive filtering module, adding the adaptive filtering module to a deep learning model, and performing building material dark-light image enhancement; the adaptive light component estimation can be performed on different feature regions in the image, the building material defects and structural features are prevented from being excessively smoothed, and the quality of dark-light image enhancement by the deep learning model is improved.

Description

Technical Field

[0001] This application relates to the field of digital image processing technology, specifically to a low-light image enhancement method, system, and medium that integrates deep learning. Background Technology

[0002] With the development of machine vision, defect detection methods based on deep learning models have been widely used in various industries (such as building material defect detection). The quality of the acquired images is crucial for improving the model's defect detection accuracy. However, in real-world industrial scenarios, factors such as ambient lighting and interference result in poor image quality in low-light conditions, severely impacting image usability and visual appeal. Low-light image enhancement aims to improve the quality of images captured under low-light conditions, reducing the effects of noise, low contrast, and color distortion on images captured at night or in low-light environments.

[0003] When ambient lighting conditions are poor, camera sensors increase ISO to capture image information more clearly. Under high ISO conditions, highly reflective areas of building materials (such as local areas of tiles or alloy plates) may be overexposed, resulting in light spots and halos. Noise in dark areas is amplified, leading to strong noise artifacts in the overall image. Image enhancement methods based on deep learning models (such as the U-Net structure) typically use an encoder-decoder structure to extract downsampled features and restore upsampled images. The feature maps extracted by the encoder at different stages mainly focus on the semantic structure and texture information of local image patches. They lack explicit modeling of illumination distribution features and noise. Furthermore, the significant differences in illumination and noise at different locations in the entire low-light image can cause subtle details and edge information to be blurred or fogged in the enhanced image, and halo artifacts may appear at the boundary between strong and dark light, affecting the quality of image enhancement. Summary of the Invention

[0004] To address the aforementioned technical problems, the purpose of this application is to provide a low-light image enhancement method, system, and medium that integrates deep learning. The specific technical solution adopted is as follows: In a first aspect, embodiments of this application provide a low-light image enhancement method incorporating deep learning, the method comprising the following steps: Acquire RGB images of building material surfaces under low-light conditions; The pixels in the RGB image are divided into different categories based on the gradient of the dark channel value; connected components are extracted for any category of pixels in the RGB image; and the first feature of each pixel is constructed based on the difference in pixel values ​​of different channels in each connected component. Analyze the complexity of texture distribution within each connected component, and determine the second feature of each connected component by combining the area of ​​each connected component. The Gaussian filtering module in the Retinex algorithm is optimized based on the first and second features of each connected component to obtain the estimated intensity of the illumination components after improvement. The Retinex algorithm and deep learning model are then combined to enhance the surface image of building materials.

[0005] In one embodiment, the process of obtaining the first feature is as follows: Calculate the difference between any two channel normalized pixel values ​​of each pixel; calculate the variance of the difference of all pixels in any two channels of each connected component; determine the first feature of each connected component based on the variance, wherein the first feature is positively correlated with the variance.

[0006] In one embodiment, determining the first feature of each connected component based on the variance specifically involves: Calculate the mean of all the variances in each connected component; The minimum value between the preset feature threshold and the mean is used as the first feature of each connected component.

[0007] In one embodiment, the process of obtaining the second feature is as follows: The ratio of the area of ​​each connected component to the average area of ​​all connected components under the same type of pixels is used as the first parameter of each connected component. Calculate the entropy value of the gray-level co-occurrence matrix of the grayscale image in the neighborhood of each pixel; determine the second parameter of each connected component based on the degree of disorder in the distribution of the entropy values ​​of all pixels in each connected component; The second feature of each connected component is determined based on the first parameter and the second parameter. The second feature is positively correlated with the first parameter and negatively correlated with the second parameter.

[0008] In one embodiment, the second parameter is the normalized value of the variance of the entropy values ​​of all pixels in each connected component.

[0009] In one embodiment, the second feature of each connected component determined based on the first parameter and the second parameter is expressed as: The first parameter is used as the numerator, the sum of the second parameter and a preset constant greater than 0 is used as the denominator, and the ratio of the numerator to the denominator is used as the second feature.

[0010] In one embodiment, the expression for optimizing the Gaussian filter in the Retinex algorithm to obtain the estimated intensity of the illumination components after improvement for each connected component is as follows: In the formula, For connected components The corresponding improved intensity of the light component estimation; For connected components The first characteristic; The intensity of the illumination component is estimated using the original Gaussian filter kernel function; For connected components The standard deviation of the corresponding Gaussian filter kernel function; For connected components The normalized value of the second feature; The initial value of the standard deviation of the Gaussian filter kernel function; , These are the preset first boundary value and the second boundary value, respectively; This is a truncation function.

[0011] In one embodiment, the process of performing surface image enhancement on building materials is as follows: In the deep learning model, the input image is filtered using an optimized Retinex algorithm. The filtered image is then fused with the original input image and input into the encoder of the U-Net network model for further processing, resulting in a low-light image enhancement model that incorporates the Retinex image filtering module. A low-light image dataset is obtained by adjusting the brightness of building material images taken under normal lighting conditions to achieve low-light performance and adding noise and blurring. The original image data taken under normal lighting conditions is used as the label data. The low-light image dataset and the labels of each image are input into the low-light image enhancement model for training. The trained low-light image enhancement model is then used to enhance real-time low-light images of building materials.

[0012] Secondly, embodiments of this application also provide a low-light image enhancement system that integrates deep learning, wherein the system stores a computer program that, when executed by a processor, implements the steps of the method described in the first aspect.

[0013] Thirdly, embodiments of this application also provide a low-light image enhancement medium incorporating deep learning, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the method described in the first aspect above.

[0014] The embodiments of this application have at least the following beneficial effects: This application acquires RGB images of building material surfaces under low-light conditions; divides pixels into different types based on dark channel value gradients; constructs the first feature of each connected region based on the degree of difference in the response of pixels in each connected region to light intensity across the RGB three channels; determines whether the local response to light intensity in the connected region is affected by changes in the reflective properties of the building material surface structure, thereby reflecting the feature complexity of the region; and adjusts the intensity estimation of the Gaussian filter illumination component in the Retinex algorithm accordingly. Then, it analyzes the relative size and local characteristics of each connected region within connected regions of the same illumination type. Based on the texture variation features, a second feature is constructed for each connected region to reflect whether the region corresponding to the connected region is a large-scale smooth region or a small-scale structurally abrupt region. Based on the second feature, the size of the Gaussian filter kernel in the Retinex algorithm is adjusted to perform different degrees of smoothing, completing the construction of the adaptive Retinex image filtering module. This module can perform adaptive illumination component estimation for different feature regions in the image, avoiding excessive smoothing of building material defects and structural features, and improving the quality of low-light image enhancement by deep learning models. It also solves the problem that traditional algorithms lack explicit modeling of illumination distribution features and noise, which affects the image enhancement quality. Attached Figure Description

[0015] Figure 1 A flowchart illustrating the steps of a low-light image enhancement method incorporating deep learning, as provided in one embodiment of this application; Figure 2 This is a schematic diagram illustrating the process of obtaining the second feature. Detailed Implementation

[0016] The following description, in conjunction with the accompanying drawings, details the specific scheme of the low-light image enhancement method, system, and medium fused with deep learning provided in this application.

[0017] Please see Figure 1 The diagram illustrates a flowchart of a low-light image enhancement method incorporating deep learning, according to an embodiment of this application. The method includes the following steps: Step S1: Acquire RGB images of building material surfaces in low-light conditions.

[0018] RGB images of building material surfaces are captured using an industrial CCD camera in low-light environments.

[0019] Step S2: Divide the pixels in the RGB image into different categories based on the gradient of the dark channel value of the pixels; extract the connected components of any category of pixels in the RGB image; and construct the first feature of each pixel based on the pixel value difference of different channels in each connected component.

[0020] Image enhancement methods based on deep learning models rely on encoder-decoder convolution operations to extract features from local image patches and restore image quality. However, they lack explicit modeling of illumination distribution features and noise, which in turn affects the quality of image enhancement.

[0021] Traditional Retinex algorithms for low-light image enhancement rely on Gaussian filtering to estimate illumination components across the entire image. The core assumption is that illumination changes in a scene exhibit strong spatial smoothness and continuity. Illumination components are then removed from the actual image to enhance and repair it. However, in images of building materials, structural variations at different locations (e.g., tile surfaces, alloy sheet surfaces) lead to different reflective properties, resulting in significant brightness abrupt changes. Furthermore, the increased sensitivity of camera sensors in low-light conditions causes substantial noise differences between locally dark and overexposed areas. These factors lead to the traditional Retinex algorithm blurring and fogging of subtle detail edges in the enhanced image, and the appearance of halo artifacts at the boundaries between strong and dark light, making it difficult to guarantee the preservation of defect features in the enhanced image.

[0022] To address the aforementioned issues, the dark channel value of each pixel in the acquired image is first calculated using the dark channel calculation method. The dark channel calculation method is an existing technology, and its specific process will not be elaborated here.

[0023] The dark channel values ​​of all pixels were statistically analyzed, and all pixels were categorized into regions with different illumination intensities based on their dark channel values. Specifically, the 33rd and 66th percentiles of the dark channel values ​​were used as thresholds to classify all pixels, resulting in three types of pixels, denoted as _____. type, type, Type: Pixels with a dark channel value below 33% belong to Pixels whose dark channel values ​​are above the 33rd percentile and below the 66th percentile belong to the category of... Type: Pixels whose dark channel value is above the 66th percentile belong to... The process involves identifying the type of pixel. Then, for each type, connected component extraction is performed on all pixels. This means that adjacent pixels belonging to the same type are merged to obtain a single connected component. Connected component extraction is a well-known technique, and the specific process will not be elaborated upon here.

[0024] Furthermore, with Taking type as an example, let the light intensity be denoted as Common types The nth connected component, where the nth is... Let the connected components be denoted as , The number of pixels in the data is denoted as To each The pixel values ​​of pixels in the R, G, and B channels are normalized. In this embodiment, the maximum value normalization method is used to normalize the pixel values, and the maximum value is 255. Then, the calculation is performed. The variance of the normalized pixel value difference of all pixels in the R, G, and B channels respectively. , , in express The variance of the difference between the normalized R-channel pixel values ​​and the normalized G-channel pixel values ​​of all pixels in the dataset. express The variance of the difference between the normalized G-channel pixel values ​​and the normalized B-channel pixel values ​​of all pixels in the dataset. express The variance of the difference between the normalized B-channel pixel values ​​and the normalized R-channel pixel values ​​of all pixels.

[0025] By quantifying the inconsistency in the response of different channels to color changes through the differences of pixels in the RGB three channels, a model is constructed. The first feature. Preferably, in this embodiment, the expression of the first feature is: In the formula, For connected components The first characteristic, For the preset feature threshold, This is a function for finding the minimum value. The denominator is used to normalize the numerator. In this embodiment, Take 0.9. In other embodiments of this application, the implementer may set the value according to the actual situation.

[0026] Because of connected components It was obtained by connecting the components after dividing them according to the dark channel values. All pixels in the light intensity are uniformly affected. If the surface structure of the building material remains continuous and uniform, the color attributes of these pixels in the RGB channels will also be affected by the same change in light intensity in a roughly consistent proportion. Conversely, if the surface of the building material itself has structural or defective features that disrupt the stable and continuous changes in the structure, the color attributes of these defective and structurally changed pixels will also change differently in response to light in the RGB channels.

[0027] So The larger the value, the more color pairs the different channels represent. Significant differences in regional illumination response and the non-uniform reflectivity of the building material surface lead to substantial variations in the reflection characteristics of different wavelengths of light across different RGB channels in that local area. In such cases, if a strong filtered illumination component is still used to estimate the intensity of the image, these local brightness variations caused by changes in local structure are easily mistaken for variations in illumination, resulting in over-smoothing of defect features and detailed structural features during image enhancement. Conversely, [the text abruptly ends here, likely due to an incomplete sentence or missing information]. The smaller the value, the better. The surface structure of the building materials in the region changes uniformly, and the corresponding changes in light intensity in each channel are uniform with minimal overall differences. In this case, the intensity is enhanced by estimating the intensity of the filtered light components of the original image.

[0028] Step S3: Analyze the complexity of texture distribution within each connected domain, and determine the second feature of each connected domain by combining the area of ​​each connected domain.

[0029] Furthermore, still based on Calculate using type as an example Average number of pixels in all connected components of the type region Then connected components The first parameter It can be represented as and The ratio. The first parameter represents the connected components. The extent to which the area (the more pixels, the larger the area) is relative to the average area of ​​all connected components of the same illumination intensity type, when Then it represents a connected component. exist A region of type 1 represents a connected component with a large area, while a region of type 2 represents a connected component with a small area. Its area is relatively small.

[0030] Furthermore, the average value method is used to connect the components. The image is converted to a grayscale image. To analyze the complexity of texture changes within the grayscale image, a 3x3 matrix window centered on each pixel is used as the neighborhood of each pixel. Then, the entropy value of the gray-level co-occurrence matrix of each pixel's neighborhood is calculated. Both grayscale image conversion and gray-level co-occurrence matrix calculations are existing technologies, and their specific processes will not be elaborated upon. The size of the pixel neighborhood can be set by the implementer according to actual conditions; this application does not impose specific limitations.

[0031] Then to The entropy values ​​of all pixels in the dataset are normalized using the maximum value normalization method. Then, based on prior knowledge, the range of the variance of the sequence with data values ​​in the range [0,1] is [0, ...]. ], then the connected component The second parameter It can be represented as The variance of the normalized entropy value of all pixels is four times the value of the second parameter to ensure that the value is between [0,1]. The maximum value normalization is a well-known technique, and the specific process will not be elaborated further.

[0032] In image analysis, the entropy value of a pixel reflects the non-uniformity and complexity of the texture in a local region of the pixel in the image. The larger the entropy value, the greater the variation in the texture features of the local region where the pixel is located. By calculating the variance of the normalized entropy values ​​of the pixels in the connected domain, the overall texture complexity of the connected domain can be represented. The larger the variance, the greater the difference in texture features of each pixel, and the greater the texture variation features in the entire connected domain.

[0033] Based on the above analysis, construct Second feature Preferably, in this embodiment, The expression can be: ,in, This is a preset parameter tuning factor, used to prevent the denominator from being 0. In this embodiment... Take 1. In other embodiments of this application, the implementer may set the value according to the actual situation. The value of .

[0034] In building material images, the structural textures, scratches, cracks, and other structural defects on the surface of objects such as tiles and alloy sheets alter the reflective properties of the surface microstructure, causing significant abrupt changes in brightness in localized areas compared to the surrounding regions. Therefore, these small-scale areas typically do not correspond to continuous, stable illumination areas formed by overall insufficient lighting, regardless of the intensity of the illumination in that area. Their characteristics are more likely to correspond to the inherent texture features of the building material surface itself or to areas of abrupt brightness changes in local pixels caused by minute cracks, pits, or other defects. When processing these connected components with significant size differences using Gaussian filtering, a differentiated processing approach should be adopted to ensure that the Retinex algorithm, in estimating illumination components, simultaneously preserves key structural information features and achieves better enhancement results.

[0035] Based on the above explanation, then when A value greater than 1 indicates that the size of the connected component is relatively large compared to other connected components of the same type; conversely, a value less than 1 indicates that the size of the connected component is relatively large compared to other connected components of the same type. A value less than 1 indicates a small spatial extent. When the overall texture variation within this connected region is significant, then... A larger value indicates significant texture changes within the connected component, which will lead to further shrinkage. Conversely, when the overall texture variation within a connected component is relatively small, then... If the value is small or even close to 0, avoid excessive shrinkage. The image undergoes relatively smooth filtering. For small-scale connected components with strong texture features, a smaller Gaussian kernel should be used during Gaussian filtering to avoid over-smoothing and loss of important information features in these regions, thus preserving more detailed information features. Conversely, for larger-scale connected components with relatively small overall pixel texture differences, which exhibit strong continuity of illumination intensity, these components typically correspond to continuous and stable regions with good or insufficient lighting conditions. Since the illumination changes of pixels within these regions are stable, a larger Gaussian kernel is used for illumination component estimation to achieve better brightness equalization and enhancement in large-scale regions.

[0036] Then, the second feature of all connected components is calculated, and the second feature of all connected components is normalized by the maximum and minimum values ​​to normalize its data size to the range [0,1].

[0037] Step S4: Optimize the Gaussian filtering module in the Retinex algorithm based on the first and second features of each connected component to obtain the estimated intensity of the improved illumination components of each connected component. Combine the Retinex algorithm and the deep learning model to enhance the surface image of the building material.

[0038] In summary, when using the Retinex algorithm for image enhancement, the Gaussian filtering module in the Retinex algorithm is improved based on the first and second features of each connected component, resulting in an adaptive Gaussian filtering module for each connected component. (The last sentence appears to be incomplete and possibly refers to a separate topic: "based on connected components...") For example, preferably, the improved Gaussian filter expression in this embodiment is: In the formula, For connected components The corresponding improved intensity of the light component estimation; For connected components The first characteristic; The intensity of the illumination component estimated by the original Gaussian filter kernel function is a well-known calculation process and will not be described in detail here. For connected components The standard deviation of the corresponding Gaussian filter kernel function; For connected components The normalized value of the second feature; The initial value of the standard deviation of the Gaussian filter kernel function can be set by the implementer according to the implementation scenario. This application does not impose any special restrictions. In this embodiment, it is used as the initial value of the standard deviation of the Gaussian filter kernel function. The size is set to 0.5; This is a truncation function, represented in this embodiment. hour , hour By using a truncation function to limit the upper and lower thresholds of the adaptive Gaussian kernel standard deviation, we can avoid memory overflow caused by excessively large values ​​when processing images in extreme cases. , These are the preset first boundary values ​​and second boundary values, which implementers can set themselves according to actual conditions. , The value of is not specifically limited in this application; in this embodiment... , Take values ​​of 0.2 and 2 respectively.

[0039] When connected components When the first feature is large, the local illumination differences in this region are caused by changes in the surface structure of the building materials or defects. In this case, the estimated intensity of the filtered illumination component in this region will be reduced to avoid destroying detailed structural and defect information. When the connected domain When the second characteristic is small, then its standard deviation This will reduce the light component obtained by small-scale smoothing during filtering, which is more gradual and can avoid the region being over-smoothed, thus preserving more detailed defect feature information.

[0040] Therefore, the above method can be used to obtain improved Gaussian filtering for each connected region under each type of pixel to complete the estimation of the illumination components of each connected region image.

[0041] The improved Gaussian filtering module obtained based on the above method is used as an adaptive Retinex image filtering module. After filtering the input image in the U-Net network model, the filtered image is fused with the original input image and then input into the encoder of the U-Net network model for processing, thus completing the construction of the U-Net low-light image enhancement model with the fused Retinex image filtering module. The model optimizer uses the Adam optimizer, and the loss function uses MSE loss. A low-light image dataset is constructed by adjusting the brightness of building material images taken under normal lighting conditions to low-light images and adding noise blurring processing to obtain the low-light image dataset. In this embodiment, the size of the low-light image dataset is 5000. In other embodiments of this application, the implementer can set the size of the low-light image dataset according to the actual situation.

[0042] The original, normally lit image data without brightness adjustment and blurring processing is used as label data. The low-light image dataset and the corresponding labels of each image in the dataset are input into the U-Net low-light image enhancement model fused with the Retinex image filtering module for model training. The training of the deep learning model is a prior art technique and will not be described in detail here.

[0043] The trained U-Net low-light image enhancement model, which incorporates the Retinex image filtering module, can be used for real-time low-light image enhancement of building materials captured in dark scenes.

[0044] The process of obtaining the second feature is illustrated in the diagram below. Figure 2 As shown.

[0045] Based on the same inventive concept as the above methods, embodiments of this application also provide a low-light image enhancement system that integrates deep learning. The system stores a computer program, which, when executed by a processor, implements the steps of any one of the methods described above for low-light image enhancement that integrates deep learning.

[0046] Based on the same inventive concept as the above methods, embodiments of this application also provide a low-light image enhancement medium that integrates deep learning, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of any one of the above-described low-light image enhancement methods that integrate deep learning.

[0047] In summary, this application provides a low-light image enhancement method incorporating deep learning. It involves acquiring RGB images of building material surfaces under low-light conditions; dividing pixels into different types based on dark channel value gradients; constructing the first feature of each connected region based on the degree of difference in the response of pixels in each connected region to light intensity across the RGB three channels; determining whether changes in the reflective properties of the building material surface structure cause differences in the response of local areas to light intensity within the connected region, thereby reflecting the feature complexity of the region; and adjusting the intensity estimation of the Gaussian filter illumination component in the Retinex algorithm. Finally, it analyzes the size of each connected region under the same illumination class. The relative size and local texture variation features of connected components are used to construct a second feature for each connected component, reflecting whether the region corresponding to the connected component is a large-scale smooth region or a small-scale structurally abrupt region. Based on the second feature, the size of the Gaussian filter kernel in the Retinex algorithm is adjusted to perform different degrees of smoothing, completing the construction of the adaptive Retinex image filtering module. This module can adaptively estimate the illumination components of different feature regions in the image, avoiding excessive smoothing of building material defects and structural features, thus improving the quality of low-light image enhancement by deep learning models. This also solves the problem that traditional algorithms lack explicit modeling of illumination distribution features and noise, which affects the image enhancement quality.

Claims

1. A low-light image enhancement method incorporating deep learning, characterized in that, The method includes the following steps: Acquire RGB images of building material surfaces under low-light conditions; The pixels in the RGB image are divided into different categories based on the gradient of the dark channel value of the pixels; connected components are extracted for any category of pixels in the RGB image; and the first feature of each connected component is constructed based on the difference in pixel values ​​of different channels of the pixels in each connected component. Analyze the complexity of texture distribution within each connected component, and determine the second feature of each connected component by combining the area of ​​each connected component. The Gaussian filtering module in the Retinex algorithm is optimized based on the first and second features of each connected component to obtain the estimated intensity of the illumination components after improvement of each connected component. Combined with the Retinex algorithm and deep learning model, the surface image of building materials is enhanced. The process of obtaining the first feature is as follows: Calculate the difference between any two channel normalized pixel values ​​of each pixel; calculate the variance of the difference between all pixels in any two channels of each connected component; determine a first feature of each connected component based on the variance, wherein the first feature is positively correlated with the variance; The determination of the first feature of each connected component based on the variance specifically involves: Calculate the mean of all the variances in each connected component; The minimum value between the preset feature threshold and the mean is used as the first feature of each connected component; The process of obtaining the second feature is as follows: The ratio of the area of ​​each connected component to the average area of ​​all connected components under the same type of pixels is used as the first parameter of each connected component. Calculate the entropy value of the gray-level co-occurrence matrix of the grayscale image in the neighborhood of each pixel; determine the second parameter of each connected component based on the degree of disorder in the distribution of the entropy values ​​of all pixels in each connected component; The second feature of each connected component is determined based on the first parameter and the second parameter. The second feature is positively correlated with the first parameter and negatively correlated with the second parameter. The second parameter is: the normalized value of the variance of the entropy values ​​of all pixels in each connected component; The second feature of each connected component is determined based on the first parameter and the second parameter, and the expression is: The first parameter is used as the numerator, the sum of the second parameter and a preset constant greater than 0 is used as the denominator, and the ratio of the numerator to the denominator is used as the second feature. The expression for optimizing the Gaussian filter in the Retinex algorithm to obtain the improved intensity estimate of the illumination components for each connected component is as follows: In the formula, For connected components The corresponding improved intensity of the light component estimation; For connected components The first characteristic; The intensity of the illumination component is estimated using the original Gaussian filter kernel function; For connected components The standard deviation of the corresponding Gaussian filter kernel function; For connected components The normalized value of the second feature; The initial value of the standard deviation of the Gaussian filter kernel function; , These are the preset first boundary value and the second boundary value, respectively; This is a truncation function; The process of enhancing the surface image of building materials is as follows: In the deep learning model, the input image is filtered using an optimized Retinex algorithm. The filtered image is then fused with the original input image and input into the encoder of the U-Net network model for further processing, resulting in a low-light image enhancement model that incorporates the Retinex image filtering module. A low-light image dataset is obtained by adjusting the brightness of building material images taken under normal lighting conditions to achieve low-light performance and adding noise and blurring. The original image data taken under normal lighting conditions is used as the label data. The low-light image dataset and the labels of each image are input into the low-light image enhancement model for training. The trained low-light image enhancement model is then used to enhance real-time low-light images of building materials.

2. A low-light image enhancement system incorporating deep learning, comprising a processor and a memory, wherein the memory stores a computer program, and the processor executes the program to implement the steps of the low-light image enhancement method incorporating deep learning as described in claim 1.

3. A low-light image enhancement medium incorporating deep learning, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the low-light image enhancement method fused with deep learning as described in claim 1.

Images