An illumination adaptive image enhancement method, system and device

Abstract

The application discloses an illumination self-adaptive image enhancement method, system and device, comprising: inputting a to-be-processed image into a trained deep curve estimation network to correspondingly output a plurality of first enhanced images; wherein the deep curve estimation network is obtained through training of taking a training image as input, taking a plurality of second enhanced images as output, calculating a first loss function on the plurality of second enhanced images, and performing parameter updating; and obtaining the intensity level corresponding to the attribute of the to-be-processed image according to a pre-set intensity level relationship, and selecting the first enhanced image corresponding to the intensity level from the plurality of first enhanced images as the optimal image enhancement result. The application can self-adaptively enhance the image, save the labor cost, and improve the stability of image enhancement and the image quality.

Description

Technical Field

[0001] This invention relates to the field of vehicle control, and more particularly to an illumination-adaptive image enhancement method, system, and device. Background Technology

[0002] With the development of society and the economy, cameras are widely used in various industries. However, due to factors such as exposure time and ambient lighting, images are often generated that are visually unappealing and weaken the effectiveness of subsequent algorithmic analysis.

[0003] The existing technology CN113256533A proposes an adaptive image enhancement method based on MSRCR, which converts the source image Iin(u,v) from RGB space to HSV space and then adjusts the image. However, the technical solution proposed by CN113256533A requires a lot of manual design of empirical values, has many parameters and depends on specific data, so the algorithm is unstable. CN105046663B, based on the characteristics of low brightness and low contrast of color images, proposes an adaptive enhancement method that simulates visual perception by studying the automatic adjustment process of pupil and photoreceptor cells to the environment. However, the dark adaptation function and the quantitative adaptation function are relatively simple and cannot guarantee the stability of the algorithm. CN111223068A proposes an adaptive non-uniform image enhancement method based on Retinex, which converts the RGB color space of the input original image into the HSI color space to obtain the chromaticity map, saturation map and brightness map, and then processes them separately. However, it requires the manual design of a series of filtering parameters, which is labor-intensive and unstable.

[0004] Existing technologies require manual parameter setting for image processing, which is costly and the parameters are not always adaptable to the image. Therefore, there is an urgent need for an illumination-adaptive image enhancement algorithm to improve image quality. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of the prior art by proposing an illumination-adaptive image enhancement method, system, and device. This invention aims to provide an illumination-adaptive image enhancement method that can save labor costs and improve the stability and quality of image enhancement.

[0006] In a first aspect, the present invention provides an illumination-adaptive image enhancement method, comprising: inputting an image to be processed into a trained depth curve estimation network, and outputting multiple first enhanced images accordingly; wherein the depth curve estimation network takes a training image as input, multiple second enhanced images as output, calculates a first loss function on the multiple second enhanced images, updates parameters, and is obtained after several training iterations;

[0007] Based on a pre-defined intensity level relationship, the intensity level corresponding to the attribute of the image to be processed is obtained, and the first enhanced image corresponding to the intensity level is selected from multiple first enhanced images as the optimal image enhancement result.

[0008] This invention employs a depth curve estimation network that can output multiple enhancements and uses intensity level relationships to filter these images. It can adaptively select enhancement images of different intensity levels based on different attributes of the input image, thus adapting to different image enhancements. This reduces the time cost of manually adjusting parameters, thereby improving the stability of using the depth curve estimation network to process different images and ultimately improving the quality of image enhancement.

[0009] Furthermore, the step of obtaining the intensity level corresponding to the attribute of the image to be processed according to a pre-set intensity level relationship, and selecting the first enhanced image corresponding to the intensity level from multiple first enhanced images as the optimal image enhancement result, specifically involves:

[0010] The average image brightness of the image to be processed is calculated, and the average image brightness is used as a variable of a preset intensity level relationship to obtain the intensity level corresponding to the attribute of the image to be processed. The first enhanced image corresponding to the intensity level is selected from multiple first enhanced images as the optimal image enhancement result.

[0011] This invention calculates the average brightness of the image to be processed, enabling it to process images with different attributes. It also estimates the intensity level of the image based on the intensity level relationship, obtaining the current intensity level of the image. Based on different intensity levels, it selects images with different degrees of enhancement, which can adapt to different image attributes and thus adaptively select images with different degrees of enhancement, thereby improving the quality of image enhancement and reducing the time cost of manual parameter tuning.

[0012] Furthermore, the step of inputting the image to be processed into the trained depth curve estimation network, which outputs multiple first enhanced images, specifically involves:

[0013] The image to be processed is input into the trained depth curve estimation network, which outputs several first feature maps. Each first enhanced image corresponds to multiple first feature maps, thus obtaining multiple first enhanced images.

[0014] This invention uses multiple enhanced images as candidate images instead of just using the output image as the enhanced image. Since the enhancement effect of the candidate images is related to the attributes of the input image, in order to obtain the optimal enhancement effect, more enhanced images should be output and the optimal enhanced image should be selected from them. This allows the enhancement to adapt to the attributes of the input image and reduces the low image enhancement quality caused by using a single output enhanced image as the final result.

[0015] The first loss function includes a spatial loss function, which is specifically calculated as follows:

[0016] The training image and the second enhanced image are sequentially divided into several image blocks, each image block being a square pixel array;

[0017] Taking each image patch as the center, the average pixel value of each image patch and its surrounding area in the training image and multiple second enhanced images are calculated sequentially to obtain multiple first spatial loss functions.

[0018] The second spatial loss function is obtained by averaging the multiple first spatial loss functions.

[0019] This invention calculates the spatial loss function of multiple enhanced images, which comprehensively considers the impact of different degrees of enhancement on parameter updates. It can adapt to images with different enhancement effects, thereby giving the depth curve network better robustness and stability, making it applicable to different image properties, and adaptively generating images with different degrees of enhancement, thus improving the quality of image enhancement.

[0020] The first loss function also includes an exposure loss function, which is specifically calculated as follows:

[0021] The difference between the average brightness of each image block in the training image and the expected brightness of multiple second enhanced images is calculated sequentially to obtain multiple first exposure loss functions.

[0022] The second exposure loss function is obtained by averaging the multiple first exposure loss functions.

[0023] Furthermore, the first loss function also includes a color consistency loss function, which is specifically calculated as follows:

[0024] The average intensity of each pair of channels in multiple second-enhanced images is calculated sequentially, resulting in multiple first color consistency loss functions.

[0025] The second color consistency loss function is obtained by averaging the multiple first color consistency loss functions.

[0026] Furthermore, the first loss function also includes a smoothing loss function, which is calculated as follows:

[0027] The horizontal and vertical gradients of several second feature maps of the training image are calculated respectively, and several first smooth loss functions are obtained accordingly.

[0028] The second smooth loss function is obtained by averaging the multiple first smooth loss functions.

[0029] Furthermore, the step of obtaining the intensity level corresponding to the attribute of the image to be processed based on the preset intensity level relationship specifically involves:

[0030] Based on the preset strength level relationship, the enhanced strength value with a decimal is obtained and rounded to obtain the strength level.

[0031] Secondly, the present invention also provides an illumination-adaptive image enhancement system, comprising:

[0032] An enhancement module is used to input the image to be processed into a trained depth curve estimation network and output multiple first enhanced images. The depth curve estimation network takes the training image as input and multiple second enhanced images as output, calculates a first loss function on the multiple second enhanced images, updates the parameters, and is obtained after several training iterations.

[0033] The determining module is used to obtain the intensity level corresponding to the attribute of the image to be processed according to a pre-set intensity level relationship, and select the first enhanced image corresponding to the intensity level from multiple first enhanced images as the optimal image enhancement result.

[0034] This invention employs a depth curve estimation network that can output multiple enhancement images and uses intensity level relationships to filter these images. It can adaptively select enhancement images of different intensity levels based on different attributes of the input image, including color intensity, illumination intensity, and exposure time. This adaptive approach to different image enhancements reduces the time cost of manually adjusting parameters, thereby improving the stability of using the depth curve estimation network to process different images and ultimately enhancing the quality of image enhancement.

[0035] Thirdly, the present invention also provides an illumination-adaptive image enhancement device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the illumination-adaptive image enhancement method as described in the first aspect. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the illumination-adaptive image enhancement process provided in an embodiment of the present invention;

[0037] Figure 2 This is a schematic diagram of the network structure of the depth curve estimation network provided in an embodiment of the present invention;

[0038] Figure 3 These are image enhancement results at different enhancement levels provided in the embodiments of the present invention;

[0039] Figure 4 This is a schematic diagram illustrating the correspondence between the image to be processed and the enhancement level provided in an embodiment of the present invention;

[0040] Figure 5 This is a schematic diagram of the structure of the illumination-adaptive image enhancement system provided in an embodiment of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] See Figure 1 This is a schematic diagram of the illumination-adaptive image enhancement process provided in an embodiment of the present invention, including steps S11 to S12, specifically:

[0043] Step S11: Input the image to be processed into the trained depth curve estimation network, and output multiple first enhanced images accordingly; wherein, the depth curve estimation network takes the training image as input, multiple second enhanced images as output, calculates the first loss function for the multiple second enhanced images, and updates the parameters, and is obtained after several training iterations.

[0044] Specifically, obtaining the trained depth curve estimation network includes: inputting the image to be processed into the trained depth curve estimation network and outputting several first feature maps, with each first enhanced image corresponding to multiple first feature maps, thus obtaining multiple first enhanced images.

[0045] It is worth noting that before using the image to be processed or the training image as input to the depth curve estimation network, the image to be processed or the training image needs to be preprocessed, including normalizing the brightness of the image to be processed or the training image so that the image pixel value is within the range of [0,1], which can speed up the convergence of the depth curve estimation network.

[0046] Preferably, the depth curve estimation network is a Zero-DCE network, which learns pixel-to-pixel mapping curves to achieve image enhancement.

[0047] For example, the present invention provides a schematic diagram of the network structure of a depth curve estimation network, see [link / reference]. Figure 2 This is a schematic diagram of the network structure of the depth curve estimation network provided in an embodiment of the present invention. In the figure, module 31 is the image to be processed as input, module 32 is the backbone network of the depth curve estimation network, and module 33 is the feature map of several channels output and the candidate image of the selected optimal image enhancement result.

[0048] It is worth noting that the first feature maps (Fine-turn Feature Maps, FFMs) with N1 channels are output. Every N2 first feature maps are called a group of feature maps (GFMs). A group of feature maps corresponds to one first enhanced image, and N3 first enhanced images can be obtained.

[0049] Preferably, the first feature map with N1 = 24 channels is output.

[0050] Preferably, each set of feature maps contains N2 = 3 first feature maps.

[0051] Preferably, N3 = 8 first enhanced images can be obtained.

[0052] For example, the present invention provides schematic diagrams of image enhancement effects at different enhancement levels, see [link / reference]. Figure 3 The figures show the image enhancement results at different enhancement levels provided in the embodiments of the present invention. In the figures, the first image is the image to be processed, and the image ah is the first enhanced image rendered with an enhancement intensity of N3 of 8 from weak to strong. The brightness of the eight first enhanced images ranges from dark to bright.

[0053] Preferably, the first enhanced images are selected from the N3 first enhanced images in descending order of strength, and N4 = 6 first enhanced images are selected as candidate images for the optimal image enhancement result.

[0054] It is worth noting that, since the enhancement strength of the N3-N4 images corresponding to the first feature maps of multiple channels extracted from the preprocessed image by the backbone network of Zero-DCE is insufficient, the first N4 enhanced images of the first enhanced image are selected as candidate images for the optimal image enhancement result, from deep to shallow rendering.

[0055] Furthermore, the depth curve estimation network updates its parameters based on the first loss function, which includes the spatial loss function, exposure loss function, color consistency loss function, and smoothing loss function.

[0056] The calculation process of the spatial loss function is as follows: the training image and the second enhanced image are divided into several image blocks in sequence, and each image block is a square pixel array; taking each image block as the center, the average pixel value of each image block of the training image and the multiple second enhanced images and the surrounding area are calculated in sequence to obtain multiple first spatial loss functions; the average value of the multiple first spatial loss functions is taken to obtain the second spatial loss function.

[0057] Preferably, each image block is 16*16 pixels.

[0058] Preferably, the first spatial loss function can be expressed as:

[0059]

[0060] Where K is the number of image patches divided from a training image, Ω(i) is the four neighboring regions centered on image patch i, including the four neighboring regions above, below, left, and right, and I and O are the average pixel values ​​of the four neighboring regions before and after local image patch enhancement, respectively.

[0061] Preferably, the second spatial loss function can be expressed as:

[0062]

[0063] Where G is the number of second-enhanced images, L spa_g Indicates corresponding to L spa The g-th GFM.

[0064] It is worth noting that, in order to obtain better candidate images, the present invention applies a spatial loss function to the GFM of all candidate second-enhanced images, thus requiring the spatial loss function to be calculated for all second-enhanced images.

[0065] Preferably, the number of candidate second enhanced images is 6.

[0066] The process of calculating the exposure loss function is as follows: the difference between the average brightness of each image block of the training image and the expected brightness of multiple second enhanced images is calculated sequentially to obtain multiple first exposure loss functions; the average of the multiple first exposure loss functions is taken to obtain the second exposure loss function.

[0067] Preferably, the first exposure loss function can be expressed as:

[0068]

[0069] Where K is the number of image patches divided from a training image, and Y... i B represents the average brightness of the i-th image patch, and B is the expected brightness of the enhanced image.

[0070] It is worth noting that the same expected brightness for the same enhanced image will cause the average brightness predicted by different enhanced images to tend to be consistent, which is detrimental to the network outputting different rendering results. Therefore, different expected brightness is set for different enhanced images, and a deeper GFM will correspond to a brighter enhanced image.

[0071] Preferably, the expected brightness settings for the multiple second enhanced images are 0.35, 0.4, 0.45, 0.5, 0.55, and 0.6, respectively.

[0072] Preferably, the second exposure loss function can be expressed as:

[0073]

[0074] Among them, L exp_g Indicates corresponding to L exp The g-th GFM.

[0075] The calculation process of the color consistency loss function is as follows: calculate the average intensity of each pair of channels in multiple second enhanced images in turn to obtain multiple first color consistency loss functions; take the average of the multiple first color consistency loss functions to obtain the second color consistency loss function.

[0076] Preferably, the first color consistency loss function can be expressed as:

[0077]

[0078] Where (p,q) represents a pair of channels, J p and J q These represent the average intensity values ​​of the p and q channels in the enhanced image (a GFM), respectively.

[0079] Preferably, the second color consistency loss function can be expressed as:

[0080]

[0081] Among them, L col_g Indicates corresponding to L col The g-th GFM.

[0082] It is worth noting that L col When applied to each GFM, color consistency constraints need to be set for images enhanced to varying degrees in order to reduce color deviation in the output image.

[0083] The calculation process of the smooth loss function is as follows: calculate the horizontal and vertical gradients of several second feature maps of the training image respectively to obtain several first smooth loss functions; take the average of the multiple first smooth loss functions to obtain the second smooth loss function.

[0084] Preferably, the second smoothing loss function can be expressed as:

[0085]

[0086] Where N is the number of GFMs, and These represent horizontal and vertical gradient operations, respectively. ξ is the RGB color space, and c is the channel.

[0087] Preferably, the number of GFMs is 8.

[0088] Preferably, the first loss function can be expressed as:

[0089] L all =L spa_all +L exp_all +W col L col +W smo L smo (8)

[0090] Among them, W col and W smo These represent the weights of the second color consistency loss function and the second smoothness loss function in the overall function, respectively.

[0091] Preferably, the weight W of the second color consistency loss function in the total function col Take 5.

[0092] Preferably, the weight W of the second smoothing loss function in the total function smo Take 200.

[0093] Preferably, 1800 training images are used as the training set for the depth curve estimation network. The network weights are updated by calculating the first loss function and using gradient descent. The depth curve estimation network is trained when the first loss function converges.

[0094] This invention calculates the spatial loss function, exposure loss function, color consistency loss function, and smoothing loss function of multiple enhanced images. It can comprehensively consider the impact of different degrees of image enhancement on parameter updates and adapt to images with different enhancement effects. This makes the depth curve network more robust and stable, applicable to different image properties, and adaptively generates images with different degrees of enhancement, thereby improving the quality of image enhancement.

[0095] Step S12: Based on the preset intensity level relationship, obtain the intensity level corresponding to the attribute of the image to be processed, and select the first enhanced image corresponding to the intensity level from multiple first enhanced images as the optimal image enhancement result.

[0096] Specifically, the average image brightness of the image to be processed is calculated, and the average image brightness is used as a variable of a preset intensity level relationship to obtain the intensity level corresponding to the attribute of the image to be processed. The first enhanced image corresponding to the intensity level is selected from multiple first enhanced images as the optimal image enhancement result. In this process, according to the preset intensity level relationship, an enhancement intensity value with a decimal is obtained and rounded to obtain the intensity level.

[0097] Preferably, the strength grade relationship can be expressed as:

[0098]

[0099] in, The image data to be processed has undergone preprocessing, with values ​​ranging from [0,1]. It is the average brightness of the image to be processed after preprocessing, with a value range of [0,1]. α and β are two conversion coefficients.

[0100] Preferably, the conversion factor α is set to -5.

[0101] Preferably, the conversion coefficient β is set to 5.

[0102] It is worth noting that since the transformation coefficient α is set to -5 and the transformation coefficient β is set to 5, the value range of D1 is [0,5]. Because a continuous value range is difficult to help the depth curve estimation network select a specific output result, D1 is rounded. See also... Figure 4 This is a schematic diagram illustrating the correspondence between the image to be processed and the enhancement level provided in an embodiment of the present invention. Since the number of candidate images N4 for the optimal image enhancement result is 6, 6 enhancement levels can be estimated, with different enhancement intervals... Different intensity levels correspond to different intensity levels; each intensity level corresponds to one of the N4 candidate images of the generated image enhancement results. The higher the intensity level, the deeper the candidate image of the image enhancement result.

[0103] This solution calculates the average brightness of the normalized image to be processed, thereby estimating the optimal enhancement level and achieving a good and stable enhancement effect. For brighter images, it can automatically select effects with a weaker rendering level, while for darker images, it can automatically select effects with a stronger rendering level, thus achieving an adaptive lighting enhancement effect.

[0104] In addition, evaluation metrics such as the Natural Image Quality Evaluator (NIQE) can be used to measure the N4 candidate images with different enhancement levels and select the best image enhancement result from them, without any limitation here.

[0105] This invention also provides a schematic diagram of an illumination-adaptive image enhancement structure, see [link / reference]. Figure 5 This is a schematic diagram of the structure of the illumination-adaptive image enhancement system provided in an embodiment of the present invention, including: enhancement module 51 and determination module 52.

[0106] The enhancement module 51 is used to input the image to be processed into the trained depth curve estimation network and output multiple first enhanced images accordingly; wherein, the depth curve estimation network takes the training image as input, multiple second enhanced images as output, calculates the first loss function on the multiple second enhanced images, and updates the parameters, and is obtained after several training iterations.

[0107] The determining module 52 is used to obtain the intensity level corresponding to the attribute of the image to be processed according to the preset intensity level relationship, and select the first enhanced image corresponding to the intensity level from multiple first enhanced images as the optimal image enhancement result.

[0108] This invention employs a depth curve estimation network that can output multiple enhancements and uses intensity level relationships to filter these images. It can adaptively select enhancement images of different intensity levels based on different attributes of the input image, thus adapting to different image enhancements. This reduces the time cost of manually adjusting parameters, thereby improving the stability of using the depth curve estimation network to process different images and ultimately improving the quality of image enhancement.

[0109] The present invention also provides an illumination-adaptive image enhancement apparatus, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the illumination-adaptive image enhancement method described above.

[0110] Those skilled in the art will understand that embodiments of this application may also include computer program products. Therefore, this application may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0111] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0112] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0113] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0114] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An illumination-adaptive image enhancement method, characterized in that, The method includes: The image to be processed is input into a trained depth curve estimation network, which outputs multiple first enhanced images. The depth curve estimation network takes the training image as input and multiple second enhanced images as output. It calculates a first loss function on the multiple second enhanced images and updates the parameters. The network is obtained after several training iterations. Based on a pre-defined intensity level relationship, the intensity level corresponding to the attribute of the image to be processed is obtained, and the first enhanced image corresponding to the intensity level is selected from multiple first enhanced images as the optimal image enhancement result, specifically: The average image brightness of the image to be processed is calculated, and the average image brightness is used as a variable of a preset intensity level relationship to obtain the intensity level corresponding to the attribute of the image to be processed. The first enhanced image corresponding to the intensity level is selected from multiple first enhanced images as the optimal image enhancement result.

2. The illumination-adaptive image enhancement method as described in claim 1, characterized in that, The process of inputting the image to be processed into the trained depth curve estimation network, which outputs multiple first enhanced images, specifically involves: The image to be processed is input into the trained depth curve estimation network, which outputs several first feature maps. Each first enhanced image corresponds to multiple first feature maps, thus obtaining multiple first enhanced images.

3. The illumination-adaptive image enhancement method as described in claim 2, characterized in that, The first loss function includes a spatial loss function, which is specifically calculated as follows: The training image and the second enhanced image are sequentially divided into several image blocks, each image block being a square pixel array; Taking each image patch as the center, the average pixel value of each image patch and its surrounding area in the training image and multiple second enhanced images are calculated sequentially to obtain multiple first spatial loss functions. The second spatial loss function is obtained by averaging the multiple first spatial loss functions.

4. The illumination-adaptive image enhancement method as described in claim 2, characterized in that, The first loss function also includes an exposure loss function, which is specifically calculated as follows: The difference between the average brightness of each image block in the training image and the expected brightness of multiple second enhanced images is calculated sequentially to obtain multiple first exposure loss functions. The second exposure loss function is obtained by averaging the multiple first exposure loss functions.

5. The illumination-adaptive image enhancement method as described in claim 2, characterized in that, The first loss function also includes a color consistency loss function, which is specifically calculated as follows: The average intensity of each pair of channels in multiple second-enhanced images is calculated sequentially, resulting in multiple first color consistency loss functions. The second color consistency loss function is obtained by averaging the multiple first color consistency loss functions.

6. The illumination-adaptive image enhancement method as described in claim 2, characterized in that, The first loss function also includes a smoothing loss function, which is calculated as follows: The horizontal and vertical gradients of several second feature maps of the training image are calculated respectively, and several first smooth loss functions are obtained accordingly. The second smooth loss function is obtained by averaging the multiple first smooth loss functions.

7. The illumination-adaptive image enhancement method as described in claim 1, characterized in that, The step of obtaining the intensity level corresponding to the attribute of the image to be processed according to the preset intensity level relationship is specifically as follows: Based on the preset strength level relationship, the enhanced strength value with a decimal is obtained and rounded to obtain the strength level.

8. An illumination-adaptive image enhancement system, characterized in that, include: An enhancement module is used to input the image to be processed into a trained depth curve estimation network and output multiple first enhanced images. The depth curve estimation network takes the training image as input and multiple second enhanced images as output, calculates a first loss function on the multiple second enhanced images, updates the parameters, and is obtained after several training iterations. The determining module is used to obtain the intensity level corresponding to the attribute of the image to be processed according to a pre-set intensity level relationship, and select the first enhanced image corresponding to the intensity level from multiple first enhanced images as the optimal image enhancement result, specifically: The average image brightness of the image to be processed is calculated, and the average image brightness is used as a variable of a preset intensity level relationship to obtain the intensity level corresponding to the attribute of the image to be processed. The first enhanced image corresponding to the intensity level is selected from multiple first enhanced images as the optimal image enhancement result.

9. An illumination-adaptive image enhancement device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the illumination-adaptive image enhancement method as described in any one of claims 1-7.

Images