RGB image low light enhancement and color temperature correction combined method, system and equipment

Abstract

The invention discloses an RGB image low-light enhancement and color temperature correction combined method, system and device. The method comprises the steps that an RGB image device is used for collecting an RGB image and preprocessing the RGB image to obtain an input data set I0; carrying out brightness discrimination on each image in the I0 by adopting a gray histogram statistical mode, and obtaining a dark image data set Idark and a bright image data set Ilight; performing illumination enhancement on the Idark based on a deep neural network enhancement model to obtain a light enhanced image data set IenLight; combining by adopting a union set mode to obtain a combined image data set I1; correcting the image color temperature of the I1 based on a color temperature adjustment method, and outputting a color temperature image data set IWB after color temperature correction; and processing the IWB by using bilinear interpolation, and recovering the original resolution of the image. According to the method, the feature extraction capability of the network is enhanced, the network stacking and parameter quantity are greatly reduced, and the operation efficiency is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and particularly to a method, system and device for combining RGB image low-light enhancement and color temperature correction. Background Art

[0002] The main difficulty in RGB image low-light enhancement lies in how to effectively increase the image brightness while avoiding problems such as noise amplification, detail loss, and color distortion. The traditional histogram equalization method, although simple and widely used, easily causes image detail blurring and color shift. The Retinex model is restricted by assumptions, has poor performance in practical applications, and is computationally complex and slow. Therefore, the method based on deep learning networks can achieve better accuracy, robustness, and speed. Further combining it with color temperature correction can enrich the system functions, and at the same time enhance the system's correction ability for low-light images and improve accuracy.

[0003] Color Constancy is the ability of the human visual system to perceive the color of an object under different lighting conditions, enabling the color of an object to remain relatively stable under different light sources. The method based on statistics depends on the assumptions and statistical information of the image, and the implementation method is simple, but it performs unstably and has limited performance under complex lighting conditions. The method based on deep learning technology can provide better robustness and accuracy and stronger adaptability. Most methods solve problems in the RAW domain (i.e., the original image data obtained by the camera sensor). However, for RAW images, the data volume is very large, which increases the computational pressure. In contrast, the computational pressure of RGB images is much smaller, and the data acquisition is also simpler and more common.

[0004] The deep learning method can automatically process low-light enhancement and color temperature adjustment. It is implemented based on deep learning and has excellent performance. This technology usually learns the mapping relationship from the input image to the target image. However, the existing technology model has a large volume and the generalization ability needs to be improved. Therefore, in practical applications, how to improve the processing efficiency when combining multiple tasks is another problem that needs to be solved. Summary of the Invention

[0005] To overcome the above technical defects in the existing RGB image low-light enhancement and color temperature adjustment processes, the present invention provides a method, system and device for combining RGB image low-light enhancement and color temperature correction, aiming to solve the problems of large scale of the technical model, low processing accuracy in some scenarios, and low computational efficiency in the existing technology, and to improve the processing accuracy while enhancing the system operation efficiency and adaptability. To achieve the above objectives, the present invention is implemented through the following technical solutions:

[0006] A method for combining RGB image low-light enhancement and color temperature correction includes the steps:

[0007] Step S1: Use an RGB image device to collect the RGB image of the scene, preprocess the RGB image, and obtain the input data set I0 of the RGB image;

[0008] Step S2: Adopt the gray histogram statistics method to perform brightness discrimination on each image in the input data set I0, and obtain the dark image data set I that needs low-light enhancement dark and the bright image data set I that does not require low-light enhancement light ;

[0009] Step S3: Based on the deep neural network enhancement model, perform light enhancement on the dark image data set I dark to obtain the light-enhanced image data set I enLight ;

[0010] Step S4: Adopt the union method to merge the light-enhanced image data set I enLight with the bright image data set I light to obtain the merged image data set I1;

[0011] Step S5: Based on the color temperature adjustment method driven by the deep network, correct the image color temperature of the merged image data set I1, and output the color temperature image data set I after color temperature correction WB ;

[0012] Step S6: Use bilinear interpolation to process the color temperature image data set I WB to restore the original resolution of the images in the color temperature image data set I WB .

[0013] Preferably, the implementation method of obtaining the input data set I0 of the RGB image in step S1 specifically includes:

[0014] Single adjustment is made to the number of horizontal pixels or the number of vertical pixels of the RGB image, while keeping the scaling ratio of the RGB image unchanged, so that both the number of horizontal pixels and the number of vertical pixels of the RGB image are limited within [R min , R max ; thereby obtaining the input data set I0 of the RGB image:

[0015] I0 = {I1, I2, I3, …, I i} Formula 1

[0016] wherein, R min represents the minimum number of pixels, R max represents the maximum number of pixels, R min is set to 512 pixels, and R max is set to 2048 pixels; I iIt represents the i-th image in the input data set I0, where i is a non-zero natural number.

[0017] Preferably, in the step S2, obtaining the dark image data set I that needs to be enhanced in low light dark The implementation manner specifically includes:

[0018] Step S21: Convert each image I in the input data set I0 i into its corresponding grayscale image I gray ;

[0019] Step S22: Statistically analyze the pixel grayscale values of the grayscale image I gray to obtain a statistical histogram;

[0020] Step S23: Set the threshold k of the pixel grayscale value d to 50, and determine the pixels with pixel grayscale values below the threshold k d as dark pixels;

[0021] Step S24: According to the statistical results of the dark pixels, determine the proportion of the pixels that need low light enhancement in each image in the input data set I0;

[0022] If the proportion of the pixels that need low light enhancement in a certain image exceeds 70%, then this image is defined as a dark image;

[0023] Step S25: Compose the dark images selected from the input data set I0 into the dark image data set I that needs to be enhanced in low light dark :

[0024] I dark ={I i |i = 1, 2,..., N dark} Formula 2

[0025] In the formula, I i represents any dark image in the dark image data set I dark , and N dark represents a non-zero natural number.

[0026] Preferably, the step S24 further includes the step: If the proportion of the pixels that need low light enhancement in a certain image does not exceed 70%, then this image is defined as a bright image;

[0027] The step S25 further includes the step: Compose the bright images not selected from the input data set I0 into the bright image data set I that does not need low light enhancement light .

[0028] Preferably, the implementation manner of obtaining the light-enhanced image data set I in the step S3 enLight specifically includes:

[0029] Step S31: Adjust the parameters of each dark image in the dark image dataset I based on the deep neural network enhancement model combining transformer and CNN dark to obtain the adjustment parameter α corresponding to each pixel position of each dark image;

[0030] Step S32: For each dark image, use the obtained adjustment parameter α at each pixel position to aggregate information to obtain the corresponding adjustment parameter map A n (x);

[0031] Step S33: Perform low-light enhancement according to the obtained adjustment parameter map A n (x), and iterate the dark image dataset I through Equation 3 dark :

[0032] E n (x) = E n-1 (x) + A n (x) * E n-1 (x) * [1 - E n-1 (x)] Equation 3

[0033] In the formula, n is the number of iterations and n = 8 is set, x represents the position of each pixel, and E n (x) is the image after the nth iteration of enhancement, and E n-1 (x) is the image after the (n - 1)th iteration of enhancement;

[0034] Step S34: Obtain the light-enhanced image dataset I enLight :

[0035] I enLight = {E(m i ; A i ) | i = 1, 2,..., N} Equation 4

[0036] In the formula, m i represents the image before enhancement, A i represents the corresponding pixel-level adjustment parameter map, and E(m i ; A i ) represents a single element in the light-enhanced image dataset I enLight .

[0037] Preferably, the implementation manner of outputting the color temperature image dataset I after color temperature correction in step S5 WB specifically includes:

[0038] Step S51: Input each image in the merged image dataset I1 into the encoder for image encoding processing, map the image after image encoding processing to a high-dimensional feature space, and generate the corresponding feature map;

[0039] Step S52: Use a multi-level CNN decoding network to decode the image encoding, map the decoded image to a color temperature value of 5500K, and output the color temperature image dataset I after color temperature correction. WB .

[0040] Preferably, the implementation manner of performing image encoding processing in step S51 specifically includes:

[0041] Step S511: Perform convolution processing on each image in the merged image dataset I1 through parallel convolutional networks respectively to obtain a set of convolutional feature maps G at different scales:

[0042] G = {G i |i = 1, 2, 3, 4} Formula 5

[0043] In the formula, Gi represents four groups of feature maps with different sizes;

[0044] Step S512: Encode each image in the merged image dataset I1 through an encoder. The encoder adopts a dual-branch structure of transformer and CNN, encodes it into a convolutional feature map, with the size becoming [W / 16, H / 16] and the number of channels being 192, and this convolutional feature map is represented by a latent value V': i for representation:

[0045] V' i = f(k i ) Formula 6

[0046] In the formula, k i represents the image in I1, and f represents feature encoding;

[0047] Step S513: For the output result (ti, ci) of the encoder, perform feature fusion through a convolutional layer to obtain V':

[0048] V' = Conv(ti, ci; θ) Formula 7

[0049] In the formula, V' represents the fused encoded feature, that is, the encoding result; ti represents the feature output by the transformer branch, ci represents the feature output by the CNN branch, and θ represents network parameters.

[0050] Preferably, the implementation manner of performing image decoding in step S52 specifically includes:

[0051] Step S521: Gradually decode the encoding result V' through a multi-level CNN decoding network, and use the set of convolutional feature maps G as the supervised feature maps of each level of the decoding network respectively;

[0052] Step S522: In each stage of the decoding process, through deconvolution operations, gradually restore the image resolution to obtain the corrected color temperature image dataset I WB :

[0053]

[0054] where g represents the deconvolution operation, V' represents the fused encoded features, G represents the feature map used to supervise the decoding output, denotes the color temperature image dataset I WB and represents a single color temperature image within the color temperature image dataset I.

[0055] An RGB image low-light enhancement and color temperature correction combined system, used to process the RGB image low-light enhancement and color temperature correction combined method. The system includes:

[0056] An RGB image terminal device, responsible for capturing the RGB image of the scene through an RGB image acquisition device and uploading the image to the cloud platform for processing;

[0057] A cloud platform, equipped with an image preprocessing module, used to receive and perform image processing on the transmitted image data;

[0058] A communication module, responsible for the transmission and communication of image data;

[0059] A low-light module, by utilizing the computing power of the cloud platform, realizes statistical analysis of the input image data, intelligently adjusts the image brightness and contrast, and selectively enhances the image details and visibility;

[0060] A color temperature correction module, used to adjust the image color temperature, correct color cast, and dynamically adjust the color temperature;

[0061] A post-processing module, used to restore the original resolution of the corrected image.

[0062] An RGB image low-light enhancement and color temperature correction combined device, used to run the RGB image low-light enhancement and color temperature correction combined system. The device includes:

[0063] A terminal device and a cloud device;

[0064] The terminal device includes a memory, a communicator, and an image collector. The memory is used to save image data, the communicator is used to communicate with the cloud device, and the image collector is used to collect image data;

[0065] The cloud device includes a cloud memory, a cloud processor, and a cloud communicator; the cloud memory is used to store cloud computer programs and image data, the cloud processor includes a central processing unit CPU and a graphics processing unit GPU; the cloud communicator is used to communicate and connect with the communicator in the terminal device.

[0066] The present invention has the following advantages and beneficial effects compared with the prior art:

[0067] By introducing a low-light module, the system of the present invention not only has the ability to perform low-light enhancement processing, but also can highlight color information while enhancing the image, thus providing a more favorable basis for subsequent color temperature adjustment, further improving the system stability and overall performance; and a dual-branch network structure based on Transformer and CNN is used as the encoder to efficiently extract image features, reduce the number of network layers, and significantly reduce the model parameter quantity and resource occupancy. In addition, the present invention can directly correct the color temperature of RGB images, eliminating the processing steps of RAW image data, simplifying the data acquisition process, and greatly reducing the calculation amount; compared with traditional color temperature correction methods, the present invention effectively solves problems such as complex operation and insufficient accuracy, and is more simple and efficient in the processing process. Description of the Drawings

[0068] Figure 1 It is a flowchart of the execution of the method of the present invention. Detailed Embodiments

[0069] Next, in combination with the drawings and specific embodiments, the present invention will be further described:

[0070] To make the purpose, technical solution and advantages of the present invention clearer and more definite, the following takes embodiments with reference to the drawings to further illustrate the present invention.

[0071] Embodiment 1: A method for combining RGB image low-light enhancement and color temperature correction includes the following steps:

[0072] Step S1: Use an RGB image device to collect an RGB image of a scene, and preprocess the RGB image to obtain an input data set I0 of the RGB image;

[0073] In step S1, the specific implementation manner of obtaining the input data set I0 of the RGB image through preprocessing includes:

[0074] Single adjustment is performed on the number of horizontal pixels or the number of vertical pixels of the RGB image, keeping the scaling ratio of the RGB image unchanged, so that both the number of horizontal pixels and the number of vertical pixels of the RGB image are limited within [R min , R max ; and then the input data set I0 of the RGB image is obtained:

[0075] I0 = {I1, I2, I3, …, I i} Equation 1

[0076] Among them, R min represents the minimum number of pixels, and R max represents the maximum number of pixels. It is set that R min is 512 pixels, and it is set that R max is 2048 pixels; I i represents the i-th image in the input data set I0, and i is a non-zero natural number.

[0077] Step S2: Use the grayscale histogram statistics method to perform brightness discrimination on each image in the input data set I0, and obtain the dark image data set I dark that requires low-light enhancement and the bright image data set I light that does not require low-light enhancement; The implementation method of obtaining the dark image data set I dark that requires low-light enhancement in this step specifically includes:

[0078] Step S21: Convert each image I i in the input data set I0 into the corresponding grayscale image I gray ;

[0079] Step S22: Statistically analyze the pixel grayscale values of the grayscale image I gray to obtain a statistical histogram. Among them, the grayscale value is an integer between 0 and 255, where 0 represents black and 255 represents white. During the statistical process, an array of size 256 is created to record the number of occurrences of each grayscale value, and the statistical result is used as an abstract representation of the image brightness.

[0080] Step S23: Set the threshold k d of the pixel grayscale value to 50, and determine the pixels with pixel grayscale values below the threshold k d as dark pixels;

[0081] Step S24: According to the statistical result of the dark pixels, determine the proportion of the pixels that need low-light enhancement in each image in the input data set I0;

[0082] If the proportion of the pixels that need low-light enhancement in a certain image exceeds 70%, then the image is defined as a dark image; if the proportion of the pixels that need low-light enhancement in a certain image does not exceed 70%, then the image is defined as a bright image;

[0083] Step S25: The dark images selected from the input data set I0 are combined to form the dark image data set I dark that requires low-light enhancement;

[0084] Idark = {I i | i = 1, 2, ..., N dark} Formula 2

[0085] Wherein, I i represents any dark image in the dark image dataset I dark and N dark represents a natural number other than 0;

[0086] The unselected bright images in the input dataset I0 are combined to form a bright image dataset I that does not require low-light enhancement light .

[0087] Step S3. Based on the deep neural network enhancement model, perform light enhancement on the dark image dataset I dark to obtain a light-enhanced image dataset I enLight ; In this step S3, the implementation method of obtaining the light-enhanced image dataset I enLight specifically includes:

[0088] Step S31. Based on the deep neural network enhancement model of transformer and CNN, adjust the parameters of each dark image in the dark image dataset I dark to obtain the adjustment parameter α corresponding to each pixel position of each dark image; and here α is a three-dimensional vector corresponding to the adjustment parameters of the three channels of the RGB image.

[0089] Step S32. For each dark image, use the obtained adjustment parameters α of each pixel position to aggregate information to obtain the corresponding adjustment parameter map A n (x). This A n (x) is a matrix corresponding to the image shape, and each element position of the matrix corresponds to the image pixel position, and the specific information is the adjustment parameter α of the pixel position; it is fitted into an adjustment parameter curve graph A n (x), and the curve of this graph can map the corresponding pixel value to the enhanced pixel value; in this way, the adjustment parameter of each pixel position of the image can be obtained, so as to flexibly adjust the image;

[0090] Step S33. Perform low-light enhancement according to the obtained adjustment parameter map A n (x), and perform iteration on the dark image dataset I dark through Formula 3:

[0091] E n (x) = E n-1 (x) + A n (x) * E n-1 (x) * [1 - E n-1 (x)] Formula 3

[0092] In the formula, n is the number of iterations and n = 8 is set, x represents the position of each pixel, and E n (x) is the image after the nth iteration enhancement, and E n-1 (x) is the image after the (n - 1)th iteration enhancement;

[0093] According to the above formula, different adjustment parameters correspond to different adjustment curves, and the input pixel values can be flexibly mapped to the target output values, thereby realizing fine enhancement of the image.

[0094] Step S34: Obtain the light-enhanced image dataset I enLight :

[0095] I enLight = {E(m i ; A i )|i = 1, 2,..., N} Formula 4

[0096] In the formula, m i represents the image before enhancement, A i represents the corresponding pixel-level adjustment parameter map, and E(m i ; A i ) represents a single element in the light-enhanced image dataset I enLight , which is the result of enhancing m i according to A i and the adjustment curve.

[0097] Step S4: Use the union method to merge the light-enhanced image dataset I enLight with the bright image dataset I light to obtain the merged image dataset I1;

[0098] Step S5: Correct the image color temperature of the merged image dataset I1 based on the color temperature adjustment method driven by the deep network, and output the color temperature image dataset I WB ;

[0099] The implementation manner of outputting the color temperature image dataset I WB after color temperature correction in the said step S5 specifically includes:

[0100] Step S51: Input each image in the merged image dataset I1 into the encoder for image encoding processing, map the image after image encoding processing to the high-dimensional feature space, and generate the corresponding feature map;

[0101] Step S52: Use the multi-level CNN decoding network to decode the image encoding, and map the decoded image to the color temperature value of 5500K, and output the color temperature image dataset I WB :

[0102] Among them, the implementation of image encoding processing in step S51 specifically includes:

[0103] Step S511: Each image in the merged image dataset I1 is respectively subjected to convolution processing through parallel convolutional networks to obtain a convolutional feature map set G at different scales:

[0104] G = {G i | i = 1, 2, 3, 4} Formula 5

[0105] Gi represents four groups of feature maps of different sizes, and the number of channels is 24, 48, 96, and 192 respectively; it is used to supervise the decoding output during the decoding process;

[0106] Step S512: Each image in the merged image dataset I1 is encoded through an encoder into a convolutional feature map, with the size becoming [W / 16, H / 16] and the number of channels being 192, and this convolutional feature map is represented by a latent value V': i For representation:

[0107] V' i = f(k i ) Formula 6

[0108] In the formula, k i represents the image in I1, and f represents feature encoding; the encoder adopted uses a dual-branch structure, and respectively adopts a transformer and a CNN structure;

[0109] Step S513: For the output result (ti, ci) of the encoder, feature fusion is performed through a convolutional layer to obtain V':

[0110] V' = Conv(ti, ci; θ) Formula 7

[0111] In the formula, V' represents the fused encoded feature, that is, the encoding result, ti represents the feature output by the transformer branch, ci represents the feature output by the CNN branch, and θ represents network parameters.

[0112] Among them, the implementation of image decoding in the above steps specifically includes:

[0113] Step S521: The encoding result V' is decoded step by step through a multi-level CNN decoding network, and the convolutional feature map set G is respectively used as the supervised feature map of each level of decoding network;

[0114] Step S522: In each level of decoding process, through deconvolution operations, the image resolution is restored step by step to obtain the corrected color temperature image dataset I WB :

[0115]

[0116]

[0117] Wherein, g represents the deconvolution operation, V' represents the fused encoded feature, and G represents the feature map used to supervise the decoded output. Denote the color temperature image dataset I WB A single color temperature image within.

[0118] Step S6: Use bilinear interpolation to process the color temperature image dataset I WB To restore the original image resolution of the color temperature image dataset I WB For each pixel in this process, it is necessary to first find the corresponding coordinates in the original image, then take the four integer coordinate pixels around it, interpolate the two pixels at the top and bottom in the horizontal direction respectively to obtain two intermediate values, and then interpolate these two intermediate values in the vertical direction to finally calculate the value of the target pixel in the new image.

[0119] Embodiment 2:

[0120] An RGB image low-light enhancement and color temperature correction combined system for processing the RGB image low-light enhancement and color temperature correction combined method. The system includes:

[0121] An RGB image terminal device responsible for capturing the RGB image of the scene through an RGB image acquisition device and uploading the image to the cloud platform for processing;

[0122] A cloud platform equipped with an image preprocessing module for receiving and performing image processing on the transmitted image data;

[0123] A communication module responsible for the transmission and communication of image data; ensuring the efficient transfer of image data between the RGB image terminal device and the cloud platform;

[0124] A low-light module that, by utilizing the computing power of the cloud platform, realizes statistical analysis of the input image data, especially for the pixel information in low-light areas; by analyzing the brightness distribution in the image, it can identify areas with insufficient illumination, and based on these analysis results, intelligently adjust the brightness and contrast of the image, selectively enhancing the details and visibility of the image, thereby improving the quality of low-light images;

[0125] A color temperature correction module that precisely adjusts the image color temperature through a neural network model to restore the natural color of the image and correct the color cast problem caused by uneven light source color temperature; it can also ensure more real and accurate color reproduction of the image by dynamically adjusting the color temperature;

[0126] A post-processing module for processing the corrected image to restore its original resolution.

[0127] Embodiment 3:

[0128] An RGB image low-light enhancement and color temperature correction combined device for running the above-mentioned RGB image low-light enhancement and color temperature correction combined system. The device includes:

[0129] A terminal device and a cloud device;

[0130] The terminal device includes a memory, a communicator, and an image collector. The memory is used to store image data. The communicator is used to communicate with the cloud device. The image collector is used to collect image data;

[0131] The cloud device includes a cloud memory, a cloud processor, and a cloud communicator, and is used to perform low-light enhancement and color temperature adjustment on the image according to the image data. The cloud memory is used to store cloud computer programs and image data. The cloud processor includes a central processing unit CPU and a graphics processing unit GPU, and is used to call the cloud computer programs and image data in the cloud memory to implement image adjustment. The cloud communicator is used to communicate and connect with the communicator in the terminal device.

[0132] The method disclosed in the present invention includes four key steps: data acquisition, image quality evaluation, light enhancement, and color temperature adjustment. First, data is obtained through an RGB acquisition device, and then the image information is analyzed and statistically processed by constructing a grayscale histogram. Based on the statistical results, the image quality is evaluated, and low-light enhancement processing is selectively performed. For low-light input, automatic light enhancement and color temperature adjustment are performed through a deep network. Due to the use of a feature extraction structure combining an attention mechanism and convolution, the feature extraction ability of the network is enhanced, the network stacking and the number of parameters are greatly reduced, and the operation efficiency is improved. When performing quality enhancement, the adjustment parameters of each pixel are first evaluated through the network, and then the image is finely adjusted at the pixel level according to the adjustment curve. After that, color temperature adjustment is performed through a network with an encoder-decoder structure to restore the image color information. In addition, the present invention also provides a processing system and device, which can realize automatic processing, is particularly suitable for image acquisition and post-processing in low-light environments, and has a wide application prospect.

Claims

1. A method for combining low-light enhancement and color temperature correction of RGB images, characterized in that, Including the steps: Step S1: Use an RGB image device to collect the RGB image of the scene, preprocess the RGB image, and obtain the input data set I0 of the RGB image; Step S2: Use the grayscale histogram statistical method to perform brightness discrimination on each image in the input data set I0, and obtain the dark image data set I that needs to be enhanced for low light dark and the bright image data set I that does not require low light enhancement light ; Step S3: Based on the deep neural network enhancement model, enhance the illumination of the dark image dataset I dark to obtain the light-enhanced image dataset I enLight ; Step S4: Use the union method to combine the light-enhanced image dataset I enLight with the bright image dataset I light to obtain the combined image dataset I1; Step S5. Correct the color temperature of the images in the merged image dataset I1 based on the color temperature adjustment method driven by the deep network, and output the color temperature image dataset I after color temperature correction WB ; Step S6. Use bilinear interpolation to process the color temperature image dataset I WB to restore the original image resolution of the color temperature image dataset I WB .

2. The RGB image low-light enhancement and color temperature correction combination method according to claim 1, wherein, The implementation manner of obtaining the input data set I0 of the RGB image in the step S1 specifically includes: Perform a single adjustment on the number of horizontal pixels or the number of vertical pixels of the RGB image, keeping the scaling ratio of the RGB image unchanged, and restricting both the number of horizontal pixels and the number of vertical pixels of the RGB image to be within [R min , R max ; and then obtain the input data set I0 of the RGB image: I0 = {I1, I2, I3, …, I i} Equation 1 Among them, R min represents the minimum number of pixels, and R max represents the maximum number of pixels. It is set that R min is 512 pixels, and it is set that R max is 2048 pixels; I i represents the i-th image in the input data set I0, where i is a non-zero natural number.

3. A method for combining low-light enhancement and color temperature correction of RGB images according to claim 1, characterized in that, In the step S2, obtain the dark image dataset I that needs low-light enhancement dark The implementation manner specifically includes: Step S21: Convert each image I in the input data set I0 i into its corresponding grayscale image I gray ; Step S22: Statistically analyze the pixel gray values of the grayscale image I gray to obtain a statistical histogram; Step S23: Set the threshold k of the pixel gray value d to 50, and determine the pixels with pixel gray values below the threshold k d as dark pixels; Step S24: According to the statistical result of the dark pixels, determine the proportion of the pixels that need low-light enhancement in each image in the input data set I0; If the proportion of the pixels that need low-light enhancement in a certain image exceeds 70%, then this image is defined as a dark image; Step S25: The dark images selected from the input data set I0 are combined to form a dark image data set I that requires low-light enhancement dark : I dark = {I i | i = 1, 2, ..., N dark} Equation 2 Wherein, I i represents any dark image in the dark image dataset I dark , and N dark represents a natural number other than 0.

4. A method for combining low-light enhancement and color temperature correction of RGB images according to claim 3, characterized in that The step S24 further includes the step: If the proportion of the pixels that need low-light enhancement in a certain image does not exceed 70%, then this image is defined as a bright image; The step S25 further includes the step of forming a bright image dataset I that does not require low-light enhancement from the unselected bright images in the input dataset I0 light .

5. A method for combining low-light enhancement and color temperature correction of RGB images according to claim 1, characterized in that The implementation of obtaining the light-enhanced image dataset I in step S3 enLight specifically includes: Step S31: Adjust the parameters of each dark image in the dark image dataset I using a deep neural network enhancement model based on a transformer and a CNN to obtain the adjustment parameter α corresponding to each pixel position of each dark image. dark For each dark image within it, obtain the adjustment parameter α corresponding to each pixel position. Step S32: For each dark image, using the adjustment parameter α obtained at each pixel position, aggregate information to obtain a corresponding adjustment parameter map A n (x); Step S33: Perform low-light enhancement according to the obtained adjusted parameter graph A n (x), and perform iteration on the dark image dataset I through Equation 3 dark as follows: E n E(x) = n-1 E(x) + A n * E(x) n-1 * [1 - E n-1 (x)] Equation 3 where n is the number of iterations and n = 8 is set, x represents the position of each pixel, and E n (x) is the image after the nth iteration enhancement, and E n-1 (x) is the image after the (n - 1)th iteration enhancement; Step S34: Obtain the light-enhanced image dataset I enLight : I enLight = {E(m i ; A i ) | i = 1, 2, ..., N} Equation 4 where m i represents the image before enhancement, A i represents the corresponding pixel-level adjustment parameter map, and E(m i ; A i ) represents a single element in the light enhancement image dataset I enLight .

6. A method for combining low-light enhancement and color temperature correction of RGB images according to claim 1, characterized in that In the step S5, output the color temperature image dataset I after color temperature correction WB The implementation manner specifically includes: Step S51: Input each image in the merged image data set I1 into the encoder for image coding processing, map the image after the image coding processing to the high-dimensional feature space, and generate the corresponding feature map; Step S52: Decode the image encoding using a multi-level CNN decoding network, map the decoded image to a color temperature value of 5500K, and output the color temperature image dataset I after color temperature correction WB .

7. A method for combining low-light enhancement and color temperature correction of RGB images according to claim 6, characterized in that The implementation manner of performing image coding processing in the step S51 specifically includes: Step S511: Perform convolution processing on each image in the merged image data set I1 through a parallel convolution network respectively to obtain the convolution feature map sets G at different scales: G = {G i | i = 1, 2, 3, 4} Formula 5 In the formula, Gi represents four groups of feature maps with different sizes; Step S512: Encode each image in the merged image dataset I1 through an encoder. The encoder adopts a dual-branch structure of transformer and CNN, and encodes it into a convolutional feature map with a size changed to [W / 16, H / 16], a number of channels of 192, and this convolutional feature map is represented by a latent value V'. i for representation: V′ i = f(k i ) Equation 6 where k i represents the image in I1, and f represents the feature encoding; Step S513: For the output result (ti, ci) of the encoder, perform feature fusion through a convolution layer to obtain V′: V′ = Conv(ti, ci; θ) Formula 7 In the formula, V′ represents the fused encoded feature, that is, the encoding result; ti represents the feature output by the transformer branch, ci represents the feature output by the CNN branch, and θ represents the network parameter.

8. A method for combining low-light enhancement and color temperature correction of RGB images according to claim 6, characterized in that, The implementation manner of performing image decoding in the step S52 specifically includes: Step S521: Decode the encoding result V′ step by step through a multi-level CNN decoding network, and use the convolution feature map sets G as the supervised feature maps of each level of decoding network respectively; Step S522: In each level of the decoding process, through deconvolution operations, gradually restore the image resolution level by level to obtain the corrected color temperature image dataset I WB : In the formula, g represents the deconvolution operation, V′ represents the fused encoded features, and G represents the feature map used to supervise the decoded output. represents the color temperature image dataset I WB a single color temperature image within.

9. An RGB image low-light enhancement and color temperature correction combined system, which is used to process the RGB image low-light enhancement and color temperature correction combined method according to any one of claims 1-8, and is characterized in that, The system includes: An RGB image terminal device, which is responsible for capturing the RGB image of the scene through an RGB image acquisition device and uploading the image to the cloud platform for processing; A cloud platform, equipped with an image preprocessing module, which is used to receive and perform image processing on the transmitted image data; A communication module, which is responsible for the transmission and communication of image data; A low-light module, which realizes the statistical analysis of the input image data by using the computing power of the cloud platform, and intelligently adjusts the image brightness and contrast, and selectively enhances the image details and visibility; A color temperature correction module, which is used to adjust the image color temperature, correct color cast, and dynamically adjust the color temperature; A post-processing module, which is used to restore the original resolution of the corrected image.

10. An RGB image low-light enhancement and color temperature correction combined device, which is used to run the RGB image low-light enhancement and color temperature correction combined system as described in claim 9, and is characterized in that, The device includes: A terminal device and a cloud device; The terminal device includes a memory, a communicator, and an image collector. The memory is used to store image data, the communicator is used to communicate with the cloud device, and the image collector is used to collect image data; The cloud device includes a cloud memory, a cloud processor, and a cloud communicator; the cloud memory is used to store cloud computer programs and image data, the cloud processor includes a central processing unit (CPU) and a graphics processing unit (GPU); the cloud communicator is used to communicate and connect with the communicator in the terminal device.

Images