Self-adaptive high dynamic range image generation method in chemical monitoring scene

Abstract

The invention discloses a self-adaptive high-dynamic-range image generation method in a chemical monitoring scene, and belongs to the technical field of high-dynamic imaging and computational imaging. Performing registration preprocessing on the acquired reference image based on a brightness consistency criterion; constructing a high dynamic range image generation network model, inputting the reference image into the model, performing feature extraction and fusion through a multi-scale network, and generating a high dynamic range image based on the fused features; performing tone mapping on the generated high dynamic range image based on a brightness consistency criterion; and comparing the generated high dynamic range image with a reference image, calculating a pixel-level loss function and multi-dimensional perception loss, and optimizing an image generation process through a back propagation training network model. According to the method, imaging details of different scales are captured through a multi-scale feature processing method, so that the generated image is better in brightness range and details, and the expressive force of the image under a complex illumination condition is ensured.

Description

Technical Field

[0001] The present invention relates to the fields of high dynamic range imaging technology and computational imaging technology, and in particular to a method for generating adaptive high dynamic range images in a chemical monitoring scenario. Background Art

[0002] In the context of the rapid development of the modern chemical industry, its production process covers every key link, from the initial stage of raw material pretreatment, to the complex and rigorous chemical reaction process, and finally to the final product packaging. These links are not only closely linked, but also place extremely high demands on visual monitoring.

[0003] The complexity and unique characteristics of chemical production environments pose significant challenges to imaging technology, particularly during high-temperature, high-pressure reactions. Traditional imaging solutions often struggle to provide sufficient visual information to support accurate operator judgment. For example, within a high-temperature, high-pressure reactor, intense light reflects and refracts multiple times through the hot steam and metal walls, creating extreme lighting conditions. This complex light field environment results in images produced by traditional imaging systems being filled with overexposed areas, making it difficult for operators to capture key dynamic changes in the material and accurately determine the start and end points of reactions, severely impacting the accuracy and efficiency of process control. Surface defect detection is particularly challenging in the preparation of fine chemical products. Due to the presence of dark, hidden areas, conventional imaging technologies are unable to fully reveal subtle defects such as scratches and color variations, resulting in significant loss of detail and posing significant challenges to product quality control. These issues not only impact product appearance, but can also compromise performance and safety, ultimately impacting value creation across the entire supply chain. Furthermore, in the storage and transportation of hazardous chemicals, once a leak occurs, the surrounding light field becomes extremely chaotic, with interference from strong light flares and shadows concealing the leak source. Under such contrast-unbalanced image conditions, traditional imaging methods find it difficult to provide clear visual information, and their ability to provide early warning of safety risks is greatly reduced, increasing the possibility of accidents and the severity of the consequences. Summary of the Invention

[0004] The purpose of the present invention is to provide an adaptive high dynamic range image generation method in a chemical monitoring scenario to address the above-mentioned problems, hoping to improve the problem that existing imaging systems cannot fully capture details under complex lighting conditions, resulting in loss of details and affecting product quality.

[0005] The technical solution adopted by the present invention is as follows: a method for generating an adaptive high dynamic range image in a chemical monitoring scenario, comprising the following steps:

[0006] Image preprocessing: performing registration preprocessing on the collected image set with multiple exposures, i.e., the reference image, based on the brightness consistency criterion;

[0007] Build an image generation model and a high dynamic range image generation network model. Input the reference image into the model, extract and fuse features through a multi-scale network, and generate a high dynamic range image based on the fused features.

[0008] Tone mapping, which performs tone mapping on high dynamic range images based on the brightness consistency criterion;

[0009] Optimize the image, compare the generated high dynamic range image with the baseline image, calculate the pixel-level loss function and multi-dimensional perceptual loss, train the network model through backpropagation, and optimize the image generation process.

[0010] Furthermore, image preprocessing specifically includes the following steps:

[0011] Perform a gamma transform on the collected multi-exposure image set Z = {Z1, Z2, Z3} to convert the image from the original low dynamic range to a linear brightness space, where Z1, Z2, and Z3 represent images with different exposure times, respectively. Z1 is the low-exposure image, Z2 is the reference image with normal exposure, and Z3 is the high-exposure image.

[0012] The gamma transform is calculated as:

[0013] X1=A·(Z1) γ ;

[0014] X2=A·(Z2) γ ;

[0015] X3=A·(Z3) γ ;

[0016] Where X1, X2, and X3 are the output images corresponding to Z1, Z2, and Z3 respectively after gamma transformation, γ is the gamma parameter in gamma transformation, and A is the constant coefficient in gamma transformation;

[0017] The exposures of images X1, X2, and X3 are registered to produce images with aligned exposures.

[0018] It should be noted that by performing a gamma transform on the multi-exposure image set Z = {Z1, Z2, Z3}, the low dynamic range image can be converted into a linear brightness space, thereby eliminating the brightness differences between images at different exposure times, allowing different exposure images to be processed under the same brightness standard, thereby improving the consistency and accuracy of image processing.

[0019] Furthermore, registering the exposure amounts of images X1, X2, and X3 specifically includes the following steps:

[0020] Adjust the exposure of image X1 to match the exposure of image X3, thereby obtaining the exposure-corrected image I1 corresponding to X1:

[0021]

[0022] Where t1 and t3 are the exposure times of image X1 and image X3 respectively, and ρ is the exposure correction coefficient;

[0023] Adjust the exposure of image X2 to match the exposure of image X3, thereby obtaining the exposure-corrected image I2 corresponding to X2:

[0024]

[0025] Where t2 is the exposure time of image X2;

[0026] The Lucas-Kanade optical flow algorithm is used to calculate the optical flow between image X3 and the exposure-corrected images I1 and I2 respectively. The calculated optical flow is used to perform bicubic interpolation on the high-exposure image X3 to obtain the exposure-corrected image I3 corresponding to image X3.

[0027] It should be noted that the exposure levels of images X1, X2, and X3 are aligned so that the data from images with different exposures are consistently adjusted at the brightness level. This helps with precise alignment in the subsequent image fusion and synthesis process, and enhances the image's performance in various complex light field environments.

[0028] Furthermore, building an image generation model includes the following steps:

[0029] A multi-scale neural network model is constructed to extract features from different scales. For each exposure-corrected image I1, I2, and I3, a multi-scale network is used for feature extraction to obtain features that characterize different details of the exposure-corrected image:

[0030]

[0031] in, The multi-scale features corresponding to the exposure-corrected images I1, I2, and I3, respectively, where N is the number of scales, N = 3;

[0032] Integrate information from multiple different scales and fuse features extracted at different scales.

[0033] It should be noted that by processing multi-scale features, details of different scales are fully captured, making the generated image clearer in brightness range and detail display.

[0034] Furthermore, fusing features extracted at different scales includes the following steps:

[0035] The features extracted at each scale are resized by bilinear interpolation.

[0036] Set F i 1 、F i 2 、F i 3 Features extracted at N scales for exposure-corrected images I1, I2, and I3, where each feature has a different size;

[0037] Through the 1×1 convolution function Conv 1×1 () Resize each feature to obtain a uniform feature size of 1×1024

[0038]

[0039] Feature fusion, element-by-element addition at the same position, fusing information from different scales:

[0040]

[0041] The fused features are mapped through the network to generate a high dynamic range image. The mapping process is implemented through the LFF grid module.

[0042] Furthermore, the LFF grid module includes multiple convolutional layers and pooling layers, and the LFF grid module maps the fused features to obtain a high dynamic range image I generated by the network. LFF :

[0043]

[0044] Among them, LFF(·) is a feature mapping network based on convolution operation, and Sigmoid(·) is the activation function used to output high-range images.

[0045] It should be noted that the LFF module combines the advantages of deep convolutional neural networks, so that the generated images not only retain the key details of low dynamic range images, but also effectively improve the contrast and brightness levels of the images, and adapt to more complex lighting conditions and scene changes.

[0046] Furthermore, despite the use of multi-scale feature processing and LFF module design, in some chemical inspection scenarios, the brightness of the image may exceed the processing range of the detection system. Therefore, further tone mapping technology is required to ensure the availability and accuracy of the image in complex environments.

[0047] Tone mapping involves the following steps:

[0048] Set a brightness reference value L key , used to control high dynamic range image I LFF Overall brightness and contrast:

[0049]

[0050] Where W is the high dynamic range image I LFF width, H is the high dynamic range image I LFF High, R (i,j) , G (i,j) 、B (i,j) They are image I LFF The values of the R channel, B channel, and G channel at the pixel point (i, j);

[0051] Calculate the brightness adjustment factor α:

[0052]

[0053] Among them, λ is the empirical coefficient, is the standard deviation of brightness;

[0054] The high dynamic range image I is adjusted by the brightness adjustment factor α LFF To perform tone mapping:

[0055]

[0056] Among them, I HDR is the mapped image after tone mapping.

[0057] It should be noted that the brightness reference value L key During the calculation process, R (i,j) , G (i,j) and B (i,j) Each color has a specific coefficient, derived from the luminance weighting formula of the ITU-R BT.709 standard (sRGB color space). This standard is based on the physiological differences in the human eye's perception of different color channels: green light (G channel) is most sensitive, red (R channel) is second, and blue (B channel) is the weakest. Therefore, the weighting method L = 0.2126R + 0.7152G + 0.0722B is set to accurately reflect the human eye's perception of the overall brightness of the image. Tone mapping reduces image distortion caused by overly bright or dark areas, making the generated image more adaptable and better able to meet the application requirements of chemical monitoring scenarios. Tone mapping can also balance the brightness, contrast, and detail of different areas in the image, making the image more layered.

[0058] In particular, compared with the traditional fixed parameter mapping method, the introduction of a dynamic adjustment factor can flexibly adjust the degree of bright part compression according to the brightness characteristics of different images, significantly improving the robustness against problems such as local overexposure and detail loss, and further enhancing the practicality and stability of the method in high-dynamic application scenarios such as chemical monitoring.

[0059] Furthermore, optimizing the image includes the following steps:

[0060] Map the image I HDR With the reference image I GT Compare and calculate the loss function to evaluate the difference between the generated image and the target image. The calculation of is as follows:

[0061]

[0062] in, is the pixel-level loss function, is a multi-dimensional perception loss function, and λ is a balance parameter used to balance the magnitude relationship between different loss functions to achieve the purpose of optimizing two objectives at the same time;

[0063] According to the results of the loss function calculation, the network is trained through the back-propagation training algorithm until the model converges.

[0064] Furthermore, the pixel-level loss function Used to measure the pixel-level difference between the generated image and the target image; multidimensional perceptual loss function Measuring the difference between the generated image and the baseline image by comparing their features in high-level feature space;

[0065] and The calculation formula is as follows:

[0066]

[0067] in, For the mapping image I HDR The value at pixel (i, j), is the reference image I GT The value at the pixel point (i, j), φ VGG is the VGG16 network pre-trained on the ImageNet dataset, φ ResNet50 It is a ResNet50 network pre-trained on the ImageNet dataset.

[0068] It should be noted that the ImageNet project is a large-scale visual database used for visual object recognition software research; the VGG16 network is a convolutional neural network model developed by the VGG team at the University of Oxford for image recognition tasks; and the ResNet50 network is a feature extraction network structure.

[0069] Furthermore, the value range of pixel point (i, j) is [0, 1].

[0070] During image preprocessing, the reference image is converted into a linear luminance space image through gamma transformation, which is set to [0,1]. When performing subsequent image exposure matching, all pixel points (i,j) are restricted to the range of [0,1] through the constraints of max(0,·) and min(·,1), so R (i,j) , G (i,j) 、B (i,j) Both are in the interval [0,1].

[0071] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0072] 1. The present invention uses a multi-scale feature processing method to capture imaging details at different scales, thereby making the generated image have better performance in brightness range and detail display, ensuring the image's expressiveness under complex lighting conditions;

[0073] 2. The present invention combines the LFF module design with a deep convolutional neural network to generate images that not only preserve key details in low dynamic range images, but also effectively improve the contrast and brightness levels of the images, thereby adapting to more complex lighting changes and scene conditions;

[0074] 3. By introducing a special multi-dimensional perceptual loss function, this paper further improves the visual quality of the generated images and enhances the adaptability of the model, enabling it to better cope with variable inputs and targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] Figure 1 is a flow chart of the method of the present invention;

[0076] Figure 2 This is a diagram of the HDR grid structure of the present invention;

[0077] Figure 3 This is a diagram of the LFF grid structure of the present invention;

[0078] Figure 4 This is a rendering of the present invention in a real chemical monitoring scenario. DETAILED DESCRIPTION

[0079] The present invention will be described in detail below with reference to the accompanying drawings.

[0080] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0081] like Figure 1 As shown, a method for generating an adaptive high dynamic range image in a chemical monitoring scenario includes the following steps:

[0082] S100: image preprocessing, performing registration preprocessing on a set of images with multiple exposures, i.e., reference images, based on a brightness consistency criterion;

[0083] S200: Build an image generation model and a high dynamic range image generation network model, input the reference image into the model, perform feature extraction and fusion through a multi-scale network, and generate a high dynamic range image based on the fused features;

[0084] S300: Tone mapping, tone mapping is performed on the generated high dynamic range image based on the brightness consistency criterion;

[0085] S400: Optimize the image, compare the generated high dynamic range image with the baseline image, calculate the pixel-level loss function and multi-dimensional perceptual loss, train the network model through backpropagation, and optimize the image generation process.

[0086] By applying multi-scale multi-frame fusion technology to image generation in chemical monitoring scenarios, and through the above-mentioned adaptive high dynamic range image generation method in chemical monitoring scenarios, the influence of strong light flare and dark shadow areas in images under complex lighting conditions can be effectively alleviated, providing chemical monitoring with clearer, more accurate and high dynamic range images, which has significant practical value and technical advantages.

[0087] Example 1

[0088] like Figure 1 As shown, in one embodiment of the present invention, the image preprocessing process is:

[0089] The multi-exposure image set Z = {Z1, Z2, Z3} collected by a camera or other device is gamma-transformed to convert the image from its original low dynamic range to a linear brightness space, where Z1, Z2, and Z3 represent images with different exposure times, Z1 being a low-exposure image, Z2 being a reference image with normal exposure, and Z3 being a high-exposure image. The gamma transform adjusts the brightness of the image to make it more linear, which is convenient for subsequent image processing.

[0090] Through gamma transformation, each image Z1, Z2, and Z3 will be calculated to obtain the corresponding output image X1, X2, and X3. The parameters in the gamma transformation are the constant coefficient A and the gamma parameter γ. Usually, A=1 to maintain brightness consistency and γ=2.2.

[0091] The gamma-transformed images X1, X2, and X3 are subjected to exposure registration processing to ensure consistency in exposure of these images.

[0092] This embodiment eliminates brightness differences between images at different exposure times through gamma transformation, so that images with different exposure times are processed under the same brightness standard, thereby improving the consistency and accuracy of image processing.

[0093] Example 2

[0094] Another embodiment of the present invention is to perform exposure registration processing on the gamma-transformed images X1, X2, and X3, which specifically includes the following steps:

[0095] Adjust the exposure of image X1 to match the exposure of image X3, thereby obtaining the exposure-corrected image I1 corresponding to X1:

[0096]

[0097] Where t1 and t3 are the exposure times of images X1 and X3 respectively, and ρ is the exposure correction coefficient, ρ = 1.0, to ensure that the adjusted image brightness meets the requirements;

[0098] Adjust the exposure of image X2 to match the exposure of image X3, thereby obtaining the exposure-corrected image I2 corresponding to X2:

[0099]

[0100] Where t2 is the exposure time of image X2;

[0101] By using the same exposure correction coefficient ρ, the exposure amounts after adjustment of X1 and X2 are consistent with the exposure amount of X3.

[0102] The Lucas-Kanade optical flow algorithm is used to calculate the optical flow between image X3 and the exposure-corrected images I1 and I2 respectively. The calculated optical flow is used to perform bicubic interpolation on the high-exposure image X3 to obtain the exposure-corrected image I3 corresponding to image X3.

[0103] By aligning the exposure of the image, the data from images with different exposures are adjusted consistently at the brightness level, which helps with precise alignment in the subsequent image fusion and synthesis process, and enhances the image's performance in various complex light field environments.

[0104] Example 3

[0105] Another embodiment of the present invention is that constructing an image generation model includes the following steps:

[0106] A multi-scale neural network model is constructed to extract features from different scales. For each exposure-corrected image I1, I2, and I3, a multi-scale network is used to extract features, thereby obtaining features that represent different details of the exposure-corrected image.

[0107] First, the multi-scale neural network function Φ is applied to the exposure correction images I1, I2, and I3 respectively. S (·), thereby obtaining feature representations at multiple scales:

[0108]

[0109] in, The multi-scale features corresponding to the exposure-corrected images I1, I2, and I3, respectively, where N is the number of scales, N = 3;

[0110] Integrate information from multiple different scales and fuse features extracted at different scales.

[0111] This embodiment fully captures details of different scales through multi-scale feature processing, making the generated image clearer in terms of brightness range and detail display.

[0112] Example 4

[0113] like Figure 3 As shown, another embodiment of the present invention is to fuse features extracted at different scales to integrate information from multiple scales, thereby improving the expressiveness of the image. The fusion of features extracted at different scales includes the following steps:

[0114] Feature resizing: The features extracted at each scale are resized through bilinear interpolation to ensure that features at different scales have the same size;

[0115] Set F i 1 、F i 2 、F i 3 Features extracted at N scales for exposure-corrected images I1, I2, and I3, where each feature has a different size;

[0116] Through the 1×1 convolution function Conv 1×1 () Resize each feature to obtain a uniform feature size of 1×1024

[0117]

[0118] Feature fusion, element-by-element addition at the same position, fuses information from different scales, and integrates feature information from different scales by element-by-element addition:

[0119]

[0120] At this point, features from different scales are added at the same location to generate a fused feature representation;

[0121] The fused features are mapped through the network to generate a high dynamic range image. The mapping process is implemented through the LFF grid module.

[0122] The LFF grid module includes multiple convolutional layers and pooling layers. The LFF grid module maps the fused features to obtain a high dynamic range image I generated by the network. LFF :

[0123]

[0124] LFF(·) is a convolution-based feature mapping network, and Sigmoid(·) is an activation function used to output high-dynamic-range images. The above process involves processing the network output using the Sigmoid activation function to generate images that meet high dynamic range requirements.

[0125] The LFF module combines the advantages of deep convolutional neural networks, so that the generated images not only retain the key details of low dynamic range images, but also effectively improve the contrast and brightness levels of the images, and adapt to more complex lighting conditions and scene changes.

[0126] Example 5

[0127] In another embodiment of the present invention, tone mapping includes the following steps:

[0128] By calculating the weighted brightness value of each pixel in the image, the average brightness of the image is obtained as the brightness reference value L key , L key For controlling high dynamic range images I LFF Overall brightness and contrast:

[0129]

[0130] Where W is the high dynamic range image I LFF width, H is the high dynamic range image I LFF High, R (i,j) , G (i,j) 、B(i,j) They are image I LFF The values of the R channel, B channel, and G channel at the pixel point (i, j);

[0131] The brightness of each pixel is calculated by the weighted coefficients of the red, green, and blue channels (0.2126, 0.7152, 0.0722), and the overall brightness of the image is finally obtained;

[0132] Calculate the brightness adjustment factor α:

[0133]

[0134] Among them, λ is the empirical coefficient, is the standard deviation of brightness; λ can be adaptively adjusted according to the discrete degree of image brightness to enhance the algorithm's adaptability to different scene lighting conditions;

[0135] The high dynamic range image I is adjusted by the brightness adjustment factor α LFF To perform tone mapping:

[0136]

[0137] Among them, I HDR is the mapped image after tone mapping.

[0138] Tone mapping is performed on high dynamic range images using a brightness reference value, making the contrast and details of the image more consistent with visual perception.

[0139] Example 6

[0140] like Figure 4 As shown, another embodiment of the present invention is that optimizing an image includes the following steps:

[0141] Map the image I HDR With the reference image I GT Compare and calculate the loss function to evaluate the difference between the generated image and the target image. The calculation is as follows:

[0142]

[0143] in, is the pixel-level loss function, is a multi-dimensional perception loss function, λ is a balancing parameter, and λ=1; it is used to balance the magnitude relationship between different loss functions to achieve the purpose of optimizing two objectives at the same time;

[0144] According to the results of the loss function calculation, the network is trained through the back-propagation training algorithm until the model converges.

[0145] Figure 4 The effect of the algorithm provided by this embodiment on real data is demonstrated. Figure 4 The three images on the left are the input multiple-exposure image set, and the one on the right is the processed HDR image. It can be seen that after the image processing method provided in this embodiment, the details in the image are more obvious, and there is no reflection or refraction of light under complex lighting conditions.

[0146] Example 7

[0147] Another embodiment of the present invention is that the pixel-level loss function Used to measure the pixel-level difference between the generated image and the target image; multidimensional perceptual loss function Measuring the difference between the generated image and the baseline image by comparing their features in high-level feature space;

[0148] and The calculation formula is as follows:

[0149]

[0150] in, For the mapping image I HDR The value at pixel (i, j), is the reference image I GT The value at pixel (i, j), φ VGG is the VGG16 network pre-trained on the ImageNet dataset, φ ResNet50 The ResNet50 network is pre-trained on the ImageNet dataset.

[0151] By introducing a special multi-dimensional perceptual loss function, the visual quality of the generated images is further improved, the adaptability of the model is enhanced, and it can better cope with changing inputs and targets.

[0152] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for generating adaptive high dynamic range images in a chemical monitoring scenario, characterized in that: The steps include: Image preprocessing: performing registration preprocessing on the collected image set with multiple exposures, i.e., the reference image, based on the brightness consistency criterion; Build an image generation model and a high dynamic range image generation network model. Input the reference image into the model, extract and fuse features through a multi-scale network, and generate a high dynamic range image based on the fused features. Tone mapping, which performs tone mapping on high dynamic range images based on the brightness consistency criterion; Optimize the image, compare the generated high dynamic range image with the baseline image, calculate the pixel-level loss function and multi-dimensional perceptual loss, train the network model through backpropagation, and optimize the image generation process.

2. The method for generating an adaptive high dynamic range image in a chemical monitoring scenario according to claim 1, characterized in that: Image preprocessing specifically includes the following steps: Perform a gamma transform on the collected multi-exposure image set Z = {Z1, Z2, Z3} to convert the image from the original low dynamic range to a linear brightness space, where Z1, Z2, and Z3 represent images with different exposure times, respectively. Z1 is the low-exposure image, Z2 is the reference image with normal exposure, and Z3 is the high-exposure image. The gamma transform is calculated as: X1=A·(Z1) γ ; X2=A·(Z2) γ ; X3=A·(Z3) γ ; Where X1, X2, and X3 are the output images corresponding to Z1, Z2, and Z3 respectively after gamma transformation, γ is the gamma parameter in gamma transformation, and A is the constant coefficient in gamma transformation; The exposures of images X1, X2, and X3 are registered to produce images with aligned exposures.

3. The method for generating an adaptive high dynamic range image in a chemical monitoring scenario according to claim 2, characterized in that: The exposure registration of images X1, X2, and X3 specifically includes the following steps: Adjust the exposure of image X1 to match the exposure of image X3, thereby obtaining the exposure-corrected image I1 corresponding to X1: Where t1 and t3 are the exposure times of image X1 and image X3 respectively, and ρ is the exposure correction coefficient; Adjust the exposure of image X2 to match the exposure of image X3, thereby obtaining the exposure-corrected image I2 corresponding to X2: Where t2 is the exposure time of image X2; The Lucas-Kanade optical flow algorithm is used to calculate the optical flow between image X3 and the exposure-corrected images I1 and I2 respectively. The calculated optical flow is used to perform bicubic interpolation on the high-exposure image X3 to obtain the exposure-corrected image I3 corresponding to image X3.

4. The method for generating an adaptive high dynamic range image in a chemical monitoring scenario according to claim 1, characterized in that: Building an image generation model involves the following steps: A multi-scale neural network model is constructed to extract features from different scales. For each exposure-corrected image I1, I2, and I3, a multi-scale network is used for feature extraction to obtain features that characterize different details of the exposure-corrected image: in, The multi-scale features corresponding to the exposure correction images I1, I2, and I3, respectively, N is the number of scales, N = 3; Φ S (·) is a multi-scale neural network function; Integrate information from multiple different scales and fuse features extracted at different scales.

5. The method for generating an adaptive high dynamic range image in a chemical monitoring scenario according to claim 4, characterized in that: The fusion of features extracted at different scales includes the following steps: The features extracted at each scale are resized by bilinear interpolation. Set F i 1 、F i 2 、F i 3 Features extracted at N scales for exposure-corrected images I1, I2, and I3, where each feature has a different size; Through the 1×1 convolution function Conv 1×1 () Resize each feature to obtain a uniform feature size of 1×1024 Feature fusion, element-by-element addition at the same position, fusing information from different scales: The fused features are mapped through the network to generate a high dynamic range image. The mapping process is implemented through the LFF grid module.

6. The method for generating an adaptive high dynamic range image in a chemical monitoring scenario according to claim 5, characterized in that: The LFF grid module includes multiple convolutional layers and pooling layers. The LFF grid module maps the fused features to obtain a high dynamic range image I generated by the network. LFF : Among them, LFF(·) is a feature mapping network based on convolution operation, and Sigmoid(·) is the activation function used to output high-range images.

7. The method for generating an adaptive high dynamic range image in a chemical monitoring scenario according to claim 1, characterized in that: Tone mapping involves the following steps: Set a brightness reference value L key , used to control high dynamic range image I LFF Overall brightness and contrast: Where W is the high dynamic range image I LFF width, H is the high dynamic range image I LFF High, R (i,j) , G (i,j) 、B (i,j) They are image I LFF The values of the R channel, B channel, and G channel at the pixel point (i, j); Calculate the brightness adjustment factor α: Among them, λ is the empirical coefficient, is the standard deviation of brightness; The high dynamic range image I is adjusted by the brightness adjustment factor α LFF To perform tone mapping: Among them, I HDR is the mapped image after tone mapping.

8. The method for generating an adaptive high dynamic range image in a chemical monitoring scenario according to claim 7, characterized in that: Optimizing an image involves the following steps: Map the image I HDR With the reference image I GT Compare and calculate the loss function to evaluate the difference between the generated image and the target image. The calculation is as follows: in, is the pixel-level loss function, is a multi-dimensional perception loss function, and λ is a balance parameter used to balance the magnitude relationship between different loss functions to achieve the purpose of optimizing two objectives at the same time; According to the results of the loss function calculation, the network is trained through the back-propagation training algorithm until the model converges.

9. The method for generating an adaptive high dynamic range image in a chemical monitoring scenario according to claim 8, characterized in that: Pixel-level loss function Used to measure the pixel-level difference between the generated image and the target image; multidimensional perceptual loss function Measuring the difference between the generated image and the baseline image by comparing their features in high-level feature space; and The calculation formula is as follows: in, For the mapping image I HDR The value at pixel (i, j), is the reference image I GT The value at the pixel point (i, j), φ VGG is the VGG16 network pre-trained on the ImageNet dataset, φ ResNet50 It is a ResNet50 network pre-trained on the ImageNet dataset.

10. The method for generating an adaptive high dynamic range image in a chemical monitoring scenario according to claim 7, characterized in that: The value range of pixel point (i, j) is [0, 1].

Images