Low-light Image Enhancement Security Method and Device Based on Deep Learning

Abstract

The present invention discloses a low-light image enhancement security method and device based on deep learning. First, a large number of low-light images covering different scenarios, lighting conditions, and object contents are collected and preprocessed. Then, a deep learning model is constructed, which includes an image light simulation module and an image enhancement module. The image light simulation module is used to simulate the influence of light sources on the image based on the low-light image data, calculate the scattering, reflection, and shadow effects after the interaction between light and the image, and generate light simulation data. The image enhancement module is used to adjust relevant parameters for enhancement processing according to the light simulation data. Then, the constructed deep learning model is trained using the preprocessed low-light image data. Finally, the trained deep learning model is used to enhance the low-light image to be processed. The present invention can significantly improve the recognition accuracy and application efficiency of the model, improve the quality of image enhancement, and the enhanced image shows faithful physical rationality and detail authenticity.

Description

Technical Field

[0001] The present invention relates to the technical field of image data processing, and more specifically, to a low-light image enhancement security method and device based on deep learning. Background Art

[0002] Low-light images refer to images taken in environments with insufficient lighting. Due to insufficient lighting, low-light images usually have problems such as insufficient feature information, loss of detail data, color distortion, and noise interference. They are easily misinterpreted, forged, and spread maliciously, posing threats and security risks. Low-light image enhancement technology can significantly improve the quality and visibility of images under low-light conditions, which is particularly important for security monitoring systems. After enhancement, it can improve the accuracy of target recognition, reduce false alarms and missed detections in security monitoring systems, and enhance the effectiveness and security reliability of the system. For the field of autonomous driving, low-light image enhancement technology can enable the vehicle's intelligent driving system to better identify pedestrians, obstacles, road signs, etc. at night or in low-light conditions, thus improving the safety of autonomous driving. Low-light image enhancement technology can enhance the image quality, making the image more real and credible, preventing misunderstandings and wrong decisions caused by image quality problems. At the same time, it reduces the risk of malicious tampering of images. The enhanced image has obvious features and is more difficult to be tampered with without being detected. Any minor change may affect the overall quality of the image. Low-light image enhancement technology can improve the visual effect of images and extract more feature detail information, enhance the real credibility of images, and better serve subsequent image processing tasks such as segmentation, detection, and tracking, which also play important roles in information security and cyberspace security.

[0003] In the field of low-light image enhancement, it covers traditional methods and methods based on deep learning. Traditional methods include histogram equalization-based methods, such as brightness-preserving dynamic histogram equalization, which avoids the reduction of visual quality by maintaining the average brightness. The intelligent contrast enhancement technology is based on the traditional histogram equalization algorithm and processes the image in partitions to enhance it; probability-based methods have also received much attention, such as multi-scale fusion dehazing models. However, traditional methods have limitations such as the need for manual adjustment of parameters, which may lead to color distortion or over-enhancement, amplification of noise, and poor processing of motion blur effects. In recent years, the development of low-light image enhancement based on deep learning has been rapid, but many low-light image enhancement security methods based on deep learning have problems such as high training costs, slow processing speeds, and limited detail processing effects. Summary of the Invention

[0004] In order to solve the problem of limited detail processing effect in the prior art, the present invention proposes a low-light image enhancement security method and device based on deep learning.

[0005] To solve the above technical problems, the first aspect of the present invention provides a low-light image enhancement security method based on deep learning, including:

[0006] Collect a large number of low-light images covering different scenarios, lighting conditions, and object contents, and preprocess the collected low-light images;

[0007] Construct a deep learning model, which includes an image light simulation module and an image enhancement module. Among them, the image light simulation module is used to simulate the influence of the light source on the image based on the low-light image data, calculate the scattering, reflection, and shadow effects after the interaction between the light and the image, and generate light simulation data. The image enhancement module is used to adjust relevant parameters for enhancement processing according to the light simulation data and output the enhanced image. The model uses a combination of mean squared error and structural similarity loss functions to guide the model to pay attention to image structure and detail preservation while increasing brightness, and introduces a perceptual loss function to ensure that the enhanced image is similar to the reference image in high-level semantic features;

[0008] Train the constructed deep learning model with the preprocessed low-light image data;

[0009] Use the trained deep learning model to enhance the low-light image to be processed.

[0010] In one implementation, preprocessing the original low-light image data includes:

[0011] Perform coordinate calibration, resolution adjustment, and format conversion on the original low-light image to generate standardized preprocessed low-light image data.

[0012] In one implementation, training the constructed deep learning model with the preprocessed low-light image data includes:

[0013] Construct a training dataset based on the preprocessed low-light image data;

[0014] Input the training dataset into the constructed deep learning model, optimize the model parameters by calculating the difference between the model prediction and the actual expectation. In each iteration, calculate the gradient according to the loss function, and then use the backpropagation algorithm to update the weights and biases of the network, where the loss function is constructed based on mean squared error, structural similarity loss function, and perceptual loss function.

[0015] In one implementation, the model uses a combination of mean squared error and structural similarity loss functions to guide the model to pay attention to image structure and detail preservation while increasing brightness, including:

[0016] Use the mean squared error loss function to measure the difference between the enhanced image and the reference normal illumination image:

[0017]

[0018] Among them, is the total number of image pixels, is the i-th pixel of the input low-brightness image, is the i-th pixel of the enhanced image output by the model, is the mean squared error loss function;

[0019] The structural similarity loss function is introduced to measure the structural similarity between two images:

[0020]

[0021] where is the input low-brightness image, is the enhanced image output by the model, is a function to measure the structural similarity between two images. The closer the value is to 1, the more similar the image structures are, the less the image distortion is, and the more similar and reliable the enhancement effect is. is the structural similarity loss function;

[0022] The perceptual loss function is:

[0023]

[0024] where is the number of feature layers, are respectively the number of channels, height, and width of the feature map of the l th layer, is the reference image, is the enhanced image output by the model, is the perceptual feature of the l th enhanced image, is the perceptual feature of the l th reference image, represents the number of channels, height, and width of the image.

[0025] In one embodiment, the image enhancement module includes a parameter adjustment sub-module, a detail enhancement sub-module, and an image output sub-module. Among them, the parameter adjustment sub-module is used to analyze the color distribution and light intensity in the image based on the light simulation data, adjust the parameters one by one to match the predetermined visual effect, and adjust the brightness level and color saturation to generate image parameter optimization data. The detail enhancement sub-module is used to adopt the image parameter optimization data to perform hierarchical separation, sharpness and contrast adjustment on the details in the image, and optimize the key visual elements in the image. The image output sub-module is used to adjust the resolution and perform color correction on the image with optimized details, and output the enhanced image.

[0026] In one embodiment, the detail enhancement sub-module is specifically used for:

[0027] Construct a detailed feature map using the gradient information of the image;

[0028] Based on the detailed feature map, a multi-scale processing method is adopted to enhance the details. Among them, for each scale of the detailed feature map in the multi-scale processing method, a detailed enhancement filter is designed for processing. During the detailed enhancement process, a noise suppression mechanism is introduced, and according to the local variance or noise estimation value of the image, the degree of detailed enhancement is adaptively adjusted;

[0029] Fuse the detailed feature maps of different scales after detailed enhancement to obtain the detailed enhancement result;

[0030] Fuse the detailed enhancement result with the image after brightness and color adjustment to obtain the final enhanced image.

[0031] In one implementation, the method further includes: performing content review on the enhanced image, and evaluating its authenticity and credibility by analyzing the feature distribution visibility, detail rationality, and distortion situation of the enhanced image. At the same time, by embedding digital watermarks and performing image content monitoring, image data with anti-counterfeiting characteristics is generated.

[0032] Based on the same inventive concept, the second aspect of the present invention provides a low-light image enhancement security device based on deep learning, including:

[0033] A data collection and preprocessing module for collecting a large number of low-light images covering different scenarios, lighting conditions, and object contents, and preprocessing the collected low-light images;

[0034] A model construction module for constructing a deep learning model. The deep learning model includes an image light simulation module and an image enhancement module. Among them, the image light simulation module is used to simulate the influence of the light source on the image based on the low-light image data, calculate the scattering, reflection, and shadow effects after the interaction between the light and the image, and generate light simulation data. The image enhancement module is used to adjust relevant parameters for enhancement processing according to the light simulation data and output the enhanced image. The image enhancement module adopts a combination of mean square error and structural similarity loss functions to guide the model to pay attention to the image structure and detail preservation while increasing the brightness, and introduces a perceptual loss function to ensure that the enhanced image is similar to the reference image in terms of high-level semantic features;

[0035] A training module for training the constructed deep learning model using the preprocessed low-light image data;

[0036] An image processing module for enhancing the low-light image to be processed using the trained deep learning model.

[0037] Based on the same inventive concept, the third aspect of the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the deep learning-based low-light image enhancement security method described in the first aspect.

[0038] Based on the same inventive concept, the fourth aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the deep learning-based low-light image enhancement security method described in the first aspect.

[0039] Compared with the prior art, the advantages and beneficial technical effects of the present invention are as follows:

[0040] The present invention provides a deep learning-based low-light image enhancement security method, constructs a deep learning-based low-light image enhancement model, and through deep learning, it can calibrate and transform the image content by processing a large amount of low-light image data, enabling the model to identify key feature structures in the image, thereby significantly improving the recognition accuracy and application efficiency of the model. Combining the ray tracing technology and the deep learning-based neural renderer to simulate the physical propagation process of light improves the quality of image enhancement. Using the low-light image enhancement algorithm of the image enhancement module, the enhanced image shows faithful physical rationality and detail authenticity, improving the effect of image enhancement. This method is applicable to general objects, provides multi-purpose applications, and provides users with a low-light image enhancement experience.

[0041] Furthermore, through artificial intelligence technology, the optimized image is subjected to trusted analysis and content review to ensure the authenticity and credibility of the image, embed digital watermarks and perform image monitoring to ensure the security and traceability of the image content, effectively integrating the accuracy, authenticity, credibility, and security of image enhancement, and comprehensively improving the reliability and performance in low-light image enhancement and use. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0043] Figure 1 It is a flowchart of the deep learning-based low-light image enhancement security method disclosed in the embodiments of the present invention;

[0044] Figure 2 It is a framework structure diagram of the deep learning-based low-light image enhancement security device disclosed in the embodiments of the present invention;

[0045] Figure 3 It is a schematic structural diagram of the computer-readable storage medium provided in the embodiment of the present invention;

[0046] Figure 4 It is a schematic structural diagram of the computer device provided in the embodiment of the present invention;

[0047] Figure 5 It is a structural diagram of the artificial intelligence security and image credibility module in the embodiment of the present invention. Detailed implementation manners

[0048] Build a model using a deep convolutional neural network and train it based on a large number of low-brightness image datasets. Combine the mean square error and structural similarity loss functions to guide the model to pay attention to image structure and detail preservation while enhancing brightness. Introduce a perceptual loss function to ensure that the enhanced image is similar to the reference image in high-level semantic features, ensuring visual effect consistency. Adopt a step-by-step optimization strategy, fine-tune the parameters according to the evaluation index, avoid over-adjustment, and improve image quality and model stability. Through multi-scale processing methods, such as Gaussian pyramid decomposition, obtain images and feature maps with different resolutions. Process the feature maps using a detail enhancement filter to enhance high-frequency detail information. Introduce a noise suppression mechanism, adaptively adjust the detail enhancement degree according to the local variance, and avoid noise amplification. It is also possible to further optimize the details by combining an adaptive strategy and other deep learning techniques. In terms of security and credibility, it is easy to cause problems such as the enhancement effect of the enhanced image not conforming to physical laws and the enhanced details being false and untrustworthy, and it is impossible to guarantee the security and reliability of the enhanced image. The low-brightness image enhancement technology still faces many challenges and requires more innovative methods to improve the quality, efficiency, and true credibility of low-brightness image enhancement.

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0050] Embodiment 1

[0051] This embodiment discloses a low-brightness image enhancement security method based on deep learning. Please refer to Figure 1 , including:

[0052] S1: Collect a large number of low-brightness images covering different scenes, lighting conditions, and object contents, and preprocess the collected low-brightness images;

[0053] S2: Construct a deep learning model, which includes an image light simulation module and an image enhancement module. Among them, the image light simulation module is used to simulate the influence of the light source on the image based on the low-light image data, calculate the scattering, reflection, and shadow effects after the interaction between the light and the image, and generate light simulation data. The image enhancement module is used to adjust relevant parameters for enhancement processing according to the light simulation data and output the enhanced image. The model uses a combination of mean square error and structural similarity loss functions to guide the model to pay attention to the image structure and detail preservation while increasing the brightness, and introduces a perceptual loss function to ensure that the enhanced image is similar to the reference image in high-level semantic features;

[0054] S3: Use the preprocessed low-light image data to train the constructed deep learning model;

[0055] S4: Use the trained deep learning model to enhance the low-light image to be processed.

[0056] Specifically, S1 is the collection and preprocessing of low-light images, S2 is the construction of the model, S3 is the training of the model, and S4 is the testing or application of the model.

[0057] Normalizing the low-light image data focuses on standardizing the image resolution and unifying its data format to ensure data consistency and compatibility for subsequent processing. The trained deep model covers key elements such as the specific type of activation function, the weight distribution inside the model, and the number of training cycles experienced. These factors together determine the performance of the model and its processing ability for low-light images. Image feature data involves the topological structure of the reconstructed grid, the grid density distribution, and the precise coordinate information of key feature points, which are crucial for accurately understanding and analyzing the image structure. Light simulation data includes parameter settings for the position of the light source, the intensity level of the light, and the angular distribution of reflected and scattered light. By accurately simulating the light effects, it provides a more realistic lighting environment reference for image enhancement. The optimized enhanced image defines the image resolution specifications, the degree of dynamic range expansion, and the relevant parameters for color correction, thereby significantly improving the visual effect of the image. Image credibility measurement mainly considers the distortion condition of the image, the rationality of details, and the visibility of feature distribution to comprehensively evaluate the authenticity and credibility of the image.

[0058] In one implementation, the method further includes S5: Conduct content review on the enhanced image, evaluate its authenticity and credibility by analyzing the visibility of feature distribution, detail rationality, and distortion of the enhanced image, and at the same time generate image data with anti-counterfeiting characteristics by embedding digital watermarks and performing image content monitoring.

[0059] In one implementation, the preprocessing of the original low-light image data includes:

[0060] Perform coordinate calibration, resolution adjustment, and format conversion on the original low-brightness image to generate standardized preprocessed low-brightness image data.

[0061] Specifically, the preprocessing process can be implemented in the following way:

[0062] Process the initial low-brightness image data, perform consistency check and alignment of coordinates, and generate aligned image data;

[0063] Detect and adjust the image pixel resolution of the aligned image data to generate calibrated image data;

[0064] Based on the calibrated image data, perform data format conversion operations to convert the image data into a unified format and generate standardized preprocessing data.

[0065] In the specific implementation process, processing the initial low-brightness image data and performing consistency check and alignment of coordinates can be implemented in the following way:

[0066] First, carry out key point detection to extract significant feature points from the image. Subsequently, these feature points will be carefully compared with the reference points in the preset standard model library. Use the least squares method to precisely optimize the key point coordinates. This process involves calculating the deviation between the data points and the reference points, and performing elimination or adjustment operations on those points whose deviations exceed the preset error threshold. This measure ensures high standards in terms of accuracy and consistency for the output aligned image. At the same time, establish a detailed error log to record the accuracy details and the correction strategies adopted during each alignment process for continuous analysis and improvement of future alignment work. Through this series of rigorous operations, the finally generated image data achieves a high degree of alignment accuracy.

[0067] Detect and adjust the image pixel resolution of the aligned image data, which can be implemented in the following way:

[0068] Based on the image data after coordinate alignment, by analyzing each pixel point, comparing its gray level difference and color with adjacent pixels, and then evaluating the necessity of resolution adjustment. By referring to the preset resolution standard, identify the difference between the current resolution of the image and the standard resolution, and select the most appropriate resolution configuration for adjustment. Use efficient interpolation or downsampling techniques to precisely adjust the image resolution. Through iterative optimization until the goal of minimizing the average resolution error is achieved. After calibration, record in detail the changes of each pixel point before and after resolution adjustment, and generate a calibration report and image data containing comparison data.

[0069] Based on the calibrated image data, perform an operation to convert the data format, convert the image data into a unified format, and generate normalized preprocessed data, which can be achieved in the following ways:

[0070] Based on the calibrated image data, evaluate the calibration situation of the image data, and select JPEG or PNG according to requirements. For the case where the image has a large gray level, high resolution, and large image file, select the JPEG format, compress the data during the conversion process, thereby optimizing the processing speed and storage requirements, and finally generate a unified image format for subsequent image processing and analysis. For the case where the image has a small gray level, low resolution, and small image file, select the PNG format, which also supports transparency operations, and perform lossless processing on the data during the conversion process, which is suitable for most cases. Finally, generate a unified image format for subsequent image processing and analysis.

[0071] In one implementation, use the preprocessed low-light image data to train the constructed deep learning model, including:

[0072] Construct a training dataset based on the preprocessed low-light image data;

[0073] Input the training dataset into the constructed deep learning model, optimize the model parameters by calculating the difference between the model prediction and the actual expectation. In each iteration, calculate the gradient according to the loss function, and then use the backpropagation algorithm to update the weights and biases of the network. Among them, the loss function is constructed based on the mean square error, structural similarity loss function, and perceptual loss function.

[0074] Specifically, S3 constructs a training dataset based on the preprocessed low-light image data and trains the model. The trained model learns the features and structures of the low-light image, and finally obtains a trained deep learning model. The specific process includes the following:

[0075] Based on the training dataset, select the network layer structure and activation function, define the input and output nodes and the number of network layers, and generate an initialized deep learning model framework; use the initialized deep learning model framework, input the low-light image data for feature extraction, analyze the structural information and texture features of the image, optimize the model parameters, and generate a feature-optimized model; then based on the feature-optimized model, perform multiple rounds of training, adjust the learning rate and loss function, and use the batch training method to optimize the generalization ability of the model, and generate a trained deep learning model. Among them, in each round of training, calculate the loss according to the difference between the current noise prediction and the actual noise, which is used to guide the model to learn how to process the noise term at different time steps. The noise term is used to reflect the noise accumulation effect based on the time step.

[0076] Among them, based on the normalized low-brightness image data, select the network layer structure and activation function, define the number of network layers and output nodes, and the specific process of generating the initial deep learning model framework is as follows:

[0077] Generate the initial deep learning model framework based on the normalized low-brightness image data. This process includes data feature analysis, network structure selection, and network parameter initialization. First, by statistically analyzing the distribution uniformity and feature density in the low-brightness image data, select an appropriate network layer structure, such as determining how many convolutional layers to use and the number of filters in each layer. For the activation function, select Leaky ReLU according to the data characteristics to increase the ability of computational non-linear processing, or Log-Sigmoid to achieve the probability distribution of the output layer. Set the width and depth of the network to achieve the expected processing speed and recognition accuracy, and generate the initial deep learning model framework.

[0078] The specific process of adopting the initial deep learning model framework, inputting the low-brightness image data for feature extraction, analyzing the structural information and texture features of the image, and optimizing the model parameters to generate the feature-optimized model is as follows:

[0079] Adopt the initial deep learning model framework and input the processed low-brightness image data (the preprocessed low-brightness image data). Extract features layer by layer through the network layer. The convolutional layer extracts the local features of the image through filters, and the pooling layer reduces the spatial dimension of the features and reduces the computational amount. To stabilize the learning process, a normalization layer may be connected after each convolutional layer. By iteratively adjusting the network parameters, optimize the model parameters including biases and weights. In this process, the loss function is used to evaluate the difference between the actual result and the predicted result. The mean absolute percentage error (MAPE) is used as the loss function, and the expression of the loss function is: , where is the number of samples, is the actual value of the th sample, is the predicted value of the th sample. Optimize the loss through the gradient descent method to improve the fitting accuracy of the model to the data and generate the feature learning model.

[0080] Based on the feature-optimized model, conduct multiple rounds of training, adjust the loss function and learning rate, and use the batch training method to optimize the generalization ability of the model. The specific process of generating the trained deep learning model is as follows:

[0081] Based on the feature optimization model, multiple rounds of training are set to adapt to the learning needs at different stages. The loss function and learning rate are adjusted to improve the training process and enhance the generalization ability of the model. For example, a relatively high learning rate may be adopted initially for rapid convergence, and the learning rate is gradually decreased as the training progresses to refine the learning effect. The batch training method is used, where a batch of data is used in each iteration to reduce memory usage and, at the same time, utilize the statistical characteristics of the batch data to stabilize the gradient estimation. As the number of training rounds increases, the loss of the model on the training set may continue to decline, but the loss on the validation set starts to rise after reaching a minimum value. At this time, the Early Stopping technique is introduced to prevent overfitting. Through repeated training, adjusting network parameters, and improving the network structure, a trained deep learning model is finally generated, which can effectively process and interpret the complex structure of low-brightness image data.

[0082] In one implementation, the model adopts a combination of mean squared error and structural similarity loss functions to guide the model to pay attention to image structure and detail preservation while enhancing brightness, including:

[0083] The mean squared error loss function is used to measure the difference between the enhanced image and the reference normal illumination image:

[0084]

[0085] where is the total number of image pixels, is the i-th pixel of the input low-brightness image, is the i-th pixel of the enhanced image output by the model, is the mean squared error loss function;

[0086] The structural similarity loss function is introduced to measure the structural similarity between two images:

[0087]

[0088] where is the input low-brightness image, is the enhanced image output by the model, is a function to measure the structural similarity between two images, and the value closer to 1 indicates that the image structures are more similar, the image distortion is smaller, and the enhancement effect is more similar and reliable, is the structural similarity loss function;

[0089] The perceptual loss function is:

[0090]

[0091] where is the number of feature layers, are respectively thel The number of channels, height, and width of the layer feature map is the reference image is the enhanced image of the type output is the l perceptual feature of the i-th enhanced image is the l perceptual feature of the i-th reference image represents the number of channels, height, and width of the image

[0092] Specifically, before training the low-light image enhancement model, a series of preparatory work needs to be carried out to build an effective training environment. First, collect a large number of low-light images covering different scenarios, lighting conditions, and object contents as the training dataset to ensure that the model can learn the image features and enhancement patterns in various low-light situations

[0093] For the model architecture, a deep convolutional neural network is adopted, which includes multiple convolutional layers, pooling layers, and fully connected layers. The convolutional layer is used to extract local features of the image. For example, by using convolutional kernels of different sizes, it captures details such as edges and textures in the low-light image. As the network depth increases, it gradually learns more abstract global features, such as object shapes and scene structures. The pooling layer plays a role in downsampling, reducing the data volume and retaining key information to improve computational efficiency. The fully connected layer is used to integrate the features extracted previously and finally generate the enhanced image

[0094] During the training process, to guide the model learning, a suitable loss function is defined. Let be the i-th pixel of the input low-light image be the i-th pixel of the image enhanced by the model. The mean squared error loss function is used to measure the difference between the enhanced image and the reference normal-light image

[0095] To better retain image details, the structural similarity (SSIM) loss function is introduced. SSIM is used to measure the structural similarity between two images. The closer its value is to 1, the more similar the image structures are, the smaller the image distortion, and the more similar and reliable the enhancement effect is. By minimizing , the model pays attention to maintaining the image structure and details while enhancing the brightness

[0096] To ensure that the enhanced image is visually similar to the reference image, a perceptual loss function is introduced. Let be the perceptual feature extracted from the image (which can be extracted by a pre-trained neural network). By minimizing the perceptual loss, the enhanced image is close to the reference image in terms of high-level semantic features, thus ensuring the consistency of visual effects

[0097] Optimize the model parameters by calculating the difference between the prediction of the computational model and the actual expectation. In each iteration, calculate the gradient according to the loss function, and then use the backpropagation algorithm to update the weights and biases of the network. For example, for the weights of the convolutional layer , the update formula is , where is the learning rate, which determines the step size of each update, is the updated weight, is the loss function of the model. Through multiple iterations of training, the model gradually learns the best mode for low-light image enhancement and can generate enhanced images with high quality, rich details, and appropriate brightness according to the input low-light image.

[0098] In one implementation, the processing flow of the image light simulation module includes:

[0099] Based on the low-light image data, optimize the BRDF model to configure the light source to simulate the effects of natural light or indoor lighting, adjust the light source position and intensity, and set multiple lighting angles and color temperatures to generate light source configuration data. The BRDF model is a bidirectional reflectance distribution function model;

[0100] Adopt the light source configuration data to analyze the interaction between light and the surface of the low-light object, calculate the reflection and scattering effects, adjust the interaction behavior of light, simulate the performance of light, and generate light image interaction data;

[0101] Through the light image interaction data, locate the shadow area between the light source and the image object, calculate the intensity and extension of the shadow, adjust the blurriness and boundary of the shadow, optimize the light-dark contrast in the image, and obtain light simulation data.

[0102] The specific process of generating light source configuration data by optimizing the BRDF model to configure the light source to simulate the effects of natural light or indoor lighting, adjusting the light source position and intensity, and setting multiple lighting angles and color temperatures is as follows:

[0103] First, use the BRDF model to simulate the effects of different types of light sources, such as natural light. According to the requirements of the model, adjust the position of the light source to match the actual environment of the scene, such as the vehicle position or pedestrian position. In addition, adjust the intensity of the light source to conform to the changes in natural light at different times. By changing the color temperature and lighting angle, such as the sunset color temperature of sunlight, further refine the light effect. The precise setting of these parameters ensures that the simulated light can realistically reflect the lighting conditions in the real world and generate detailed light source configuration data.

[0104] The optimized BRDF model follows the formula:

[0105]

[0106] Calculate the relationship between incident light and reflected light when incident from multiple angles, and simulate the natural lighting effect;

[0107] Among them, the function represents the bidirectional reflectance distribution function, which is used to calculate the light reflectance of the image in the incident light direction and the reflected light direction The light reflectance under the condition of is the incident light direction, is the reflected light direction, is the incident angle, is the viewing angle, is the specular highlight attenuation factor, which adjusts the intensity of specular reflection according to the viewing angle so that the highlight effect weakens as the viewing angle increases, is the light intensity in the unit incident direction, indicating the light energy received per unit area in the incident direction ; is the light radiation intensity in the unit reflection direction, indicating the light brightness per unit area in the reflection direction ; is the diffuse reflection coefficient, indicating the proportion of the diffuse reflection part, is the cosine value of the angle, which is used to calculate the effective component of the light incident on the surface.

[0108] It should be noted that for a three-dimensional image, it can be understood as being mapped into a three-dimensional object, and the material property material needs to be considered; the present invention processes two-dimensional images, focusing on the quality processing of the picture itself and not mapping to three dimensions. Therefore, the material does not need to be considered when optimizing the BRDF model.

[0109] In the specific implementation process, the execution process is as follows:

[0110] Determine the angle parameters and :

[0111] (Incident angle) is determined by the angle between the incident light direction and the normal. In actual calculation, can be obtained by the arccosine of the dot product of the light source direction vector and the normal vector: It In is the unit normal vector, is the unit incident light vector.

[0112] (Viewing angle) is determined by the angle between the reflected light direction and the observer direction. Similar to , It can be obtained by taking the arccosine of the dot product of the reflected light direction vector and the viewing direction vector: where is the reflection vector, is the viewing vector.

[0113] Specular reflection coefficient and diffuse reflection coefficient Determination:

[0114] These coefficients are usually preset according to the optical and physical properties of specific objects in the image, and can be obtained from a database or determined through experimental measurements. They reflect the ability of different materials to absorb and reflect light.

[0115] Illumination intensity and light radiation intensity Calculation:

[0116] is the illumination intensity received by the object surface in the given incident direction and is usually obtained by measuring the actual illumination environment or through ray tracing. is the light radiation intensity reflected from the object surface in the given reflection direction and is usually simulated through ray tracing algorithms in rendering software.

[0117] The final BRDF value is calculated by inserting all the parameter values obtained from the above calculations into the formula. This is to comprehensively consider the influence of light angles and light intensities on reflection characteristics. The calculated result is used to accurately simulate and enhance the visual appearance of objects in low-light images under different lighting conditions. This includes how light is absorbed, scattered, and reflected by the object surface, enabling the generation of realistic enhanced images, which are widely used in security monitoring, autonomous driving, target recognition, and any applications that require precise lighting.

[0118] The specific process of using light source configuration data to analyze the interaction between light and the surface of low-light objects, calculate reflection and scattering effects, adjust the interaction behavior of light, simulate the performance of light, and generate light image interaction data is as follows:

[0119] Analyze the interaction mechanism between light and the surface of the image object. This module first uses the bidirectional reflectance distribution function model to accurately calculate the scattering and reflection situations when the light emitted from the light source touches the surface of the image object.

[0120] Use light source configuration data to analyze the interaction mechanism between light and the surface of the image object. This module first uses the bidirectional reflectance distribution function model to calculate the scattering and reflection effects when the light emitted from the light source encounters the surface of the image object. Specifically, when calculating the reflection intensity, the incident angle and the reflection angle , and its calculation formula is: Reflection intensity = Light source intensity ×BRDF(θ, )×cos(θ) , where the light source intensity represents the initial intensity of the light emitted by the light source, BRDF(θ, ) is the reflection coefficient, cos(θ) represents the effect of the incident angle on the light intensity. The scattering effect depends on the scattering coefficient and roughness of the object surface, and its calculation formula is: Scattered light intensity = Light source intensity × Scattering coefficient × (1 / 2π), where the scattering coefficient represents the scattering ability of the object surface, and (1 / 2π) is the normalization coefficient for omnidirectional scattering. Through the above series of calculations, this module can effectively simulate the specific performance of light in various environments. The generated light environment interaction data can not only accurately present the physical interaction between light and objects, but also well adapt to the actual needs of various visual scenes.

[0121] The specific process of obtaining the light simulation data by locating the shadow area between the light source and the image object through the light image interaction data, calculating the intensity and extension of the shadow, adjusting the blur and boundary of the shadow, and optimizing the light-dark contrast in the image is as follows:

[0122] Locate the shadow area generated by the obstruction between the light source and the object through the interaction data between the light and the object in the image. Use ray tracing technology to calculate the straight-line paths from the light source to each part of the object, and detect whether these paths are blocked by other objects to determine the shadow formation area. The intensity and extension of the shadow are calculated according to the distance and angle of the light source, and the "shadow attenuation factor" is used to adjust the blur of the shadow edge. This factor is adjusted according to the function of the distance from the light source to the shadow area, and the specific calculation formula is: = Shadow intensity / Base light intensity, where is the distance from the light source to the shadow generation point, is the attenuation coefficient. Thus, the light-dark contrast in the image is optimized, and accurate light simulation data is finally obtained.

[0123] In one implementation, the image enhancement module includes a parameter adjustment sub-module, a detail enhancement sub-module, and an image output sub-module. Among them, the parameter adjustment sub-module is used to analyze the color distribution and light intensity in the image based on the light simulation data, adjust the parameters one by one to match the predetermined visual effect, and adjust the brightness level and color saturation to generate image parameter optimization data. The detail enhancement sub-module is used to adopt the image parameter optimization data to perform hierarchical separation, sharpness and contrast adjustment on the details in the image, and optimize the key visual elements in the image. The image output sub-module is used to adjust the resolution and perform color correction on the detail-optimized image, and output the enhanced image.

[0124] In the specific implementation process, the parameter adjustment sub-module is based on an in-depth analysis of the characteristics of low-brightness images. First, perform statistical analysis on the light intensity and color distribution of the input low-brightness image, and calculate statistical quantities such as the average brightness of the image, the brightness histogram, and the mean and variance of each color channel (such as the RGB channel). Based on these statistical data, adjust the brightness and color saturation of the image. For brightness adjustment, according to the difference between the average brightness of the image and the target brightness, improve the overall brightness through linear or non-linear transformation. For example, let be the pixel value of the image at the coordinate , and the adjusted brightness , where is the brightness adjustment factor calculated according to the average brightness. If the average brightness is low, then is a positive value to increase the brightness. For color saturation adjustment, based on the statistical information of the color channels, enhance the channels with low saturation. For example, if the mean of the red channel is low and the variance is small, the color saturation can be increased by increasing the weight of the red channel. Let the adjusted pixel value be , where is the saturation adjustment coefficient, refers to the values of the red channel, green channel, and blue channel, is the value of the adjusted red channel.

[0125] In one implementation, the detail enhancement sub-module is specifically used for:

[0126] Construct a detail feature map using the gradient information of the image;

[0127] Based on the detail feature map, adopt a multi-scale processing method to enhance the details. Among them, for each scale of the detail feature map in the multi-scale processing method, design a detail enhancement filter for processing, introduce a noise suppression mechanism during the detail enhancement process, and adaptively adjust the degree of detail enhancement according to the local variance or noise estimate value of the image;

[0128] Fuse the detail-enhanced feature maps of different scales to obtain the detail enhancement result;

[0129] Fuse the detail enhancement result with the image after brightness and color adjustment to obtain the final enhanced image.

[0130] In the specific implementation process, the detail enhancement sub-module uses image parameter optimization data to perform hierarchical separation of details in the image, sharpness and contrast adjustment, and optimize key visual elements in the image, including edges and textures. The specific process of obtaining the detail-optimized image is as follows:

[0131] The detail enhancement sub-module aims to further improve the clarity and richness of the image. First, use the gradient information of the image to construct a detail feature map. Let be the original low-brightness image, and calculate its horizontal and vertical gradients and , for example, by calculating using the Sobel operator. Then, take the gradient magnitude as the detail feature map, which highlights the detail information such as edges and textures in the image.

[0132] Based on the detail feature map, a multi-scale processing method is used to enhance the details. Perform downsampling and upsampling operations on the original low-brightness image and the detail feature map at different scales to obtain multiple image and feature map representations with different resolutions. For example, obtain the low-resolution image ( represents different scales) and the corresponding detail feature map .

[0133] For each scale of the detail feature map , design a detail enhancement filter for processing. A simple method is to use the Laplace filter , which is clearer and sharper.

[0134] To avoid noise amplification, a noise suppression mechanism is introduced during the detail enhancement process. According to the local variance or noise estimate value of the image, adaptively adjust the degree of detail enhancement. For example, in areas with high noise, reduce the weight of the detail enhancement filter and the amplitude of detail enhancement to maintain the smoothness of the image while enhancing the details.

[0135] Fuse the detail-enhanced feature maps at different scales to obtain the final detail enhancement result. The fusion process can use the weighted summation method, and different weights are assigned according to the importance of each scale feature map. For example, a larger weight is assigned to the feature map at a lower scale (higher resolution) because it contains more detail information, while a smaller weight is assigned to the feature map at a higher scale (lower resolution) to maintain overall consistency. Let the fused detail-enhanced image be , then , where is the weight of the th scale, and , is the feature map of the th scale after detail enhancement, is the total number of scales.

[0136] Finally, fuse the detail enhancement result with the image after brightness and color adjustment to obtain the final enhanced image. Let the image after brightness and color adjustment be , then the final enhanced image , where is the weight parameter for controlling the degree of detail enhancement. Through reasonable adjustment, the effects of brightness, color, and detail enhancement can be balanced to obtain a low-brightness image enhancement result with good visual effects and rich details.

[0137] The specific process of adjusting the resolution and performing color correction on the image optimized by the above-mentioned details and outputting the enhanced image is as follows:

[0138] Perform final resolution adjustment and color correction on the image optimized by details to adapt to different output requirements and display devices. The color correction is based on the standard configuration in the color management system, adjusts each color channel, and thus achieves the color performance in the real world. The resolution adjustment is based on the requirements of the output medium, and adjusts the image resolution through sampling and resampling techniques to ensure that the image maintains the best quality at various sizes. The output optimized and rendered image thus has high-quality clarity and color accuracy.

[0139] In one implementation, the GPU cost is reduced by using Retinex-Net, an image network for low-light image enhancement. By optimizing the image processing flow in the enhancement task, Retinex-Net is end-to-end trainable and can directly learn the output enhanced image from the input low-light image without complex manual feature engineering and adjustment of intermediate steps. Retinex-Net adopts a multi-scale approach to estimate the brightness of the image and can analyze and process the illumination information of the image at different scales. Images taken in low-light environments often contain more noise. Retinex-Net can learn to predict the noise level in the image and denoise the reflectance image through a reflectance component denoising network, thus effectively reducing the noise in the image. Retinex-Net mainly includes Decom-Net (decomposition network) and Enhance-Net (enhancement network). Decom-Net can decompose the input low-light image into a reflectance image and an illumination image. The reflectance image contains the inherent attributes of the image and is independent of the illumination conditions, while the illumination image reflects the illumination information in the image. Through this decomposition, the composition of the image can be better understood, providing a basis for subsequent enhancement processing. Enhance-Net can enhance the illumination component by using methods such as adaptive mapping curves based on the estimated illumination image, thereby increasing the overall brightness of the image, improving the visibility of low-light images, making the details in the image more clearly visible, and being able to better restore the color information in the image and avoid problems such as color distortion. This model is applicable to fields that require high credibility and good effects in low-light image enhancement, such as security monitoring, autonomous driving, target recognition, etc. Compared with most deep learning models for low-light image enhancement, the Retinex-Net model is more efficient in resource use. This model reduces the dependence on high-performance computing resources, significantly reducing costs and energy consumption.

[0140] In one implementation, the method further includes: performing content review on the enhanced image, and evaluating its authenticity and credibility by analyzing the visibility of the feature distribution, the rationality of the details, and the distortion situation of the enhanced image. At the same time, by embedding digital watermarks and performing image content monitoring, image data with anti-counterfeiting characteristics is generated.

[0141] Specifically, the specific implementation process of the artificial intelligence security and image credibility module is as follows:

[0142] Based on the enhanced image, screen the feature information of the image, identify the visibility of the feature distribution, the rationality of the details, and the distortion situation, and obtain an analysis result to measure the true credibility of the enhanced image.

[0143] According to the analysis results, check the feature information marked in the image, conduct a review based on pre-trained data or a preset artificial intelligence model, delete the images that do not meet the standards, and obtain the review confirmation results.

[0144] Based on the review confirmation results, insert a digital watermark into the image, set up a monitoring program to track the distribution and use of the image, ensure that the source and distribution records of the image are queryable, and obtain anti-counterfeiting image data.

[0145] Among them, screening the feature information of the image, identifying the visibility of the feature distribution, the rationality of details, and the distortion situation, generating an analysis result, and measuring and enhancing the authenticity and credibility of the image can be achieved through the following process:

[0146] By using multi-level image scanning technology to analyze each pixel one by one, the system can comprehensively cover and scan the entire image data, and screen the feature information of each area. In this process, the data of each area is accurately analyzed and recorded, and the system uses advanced algorithms to identify the visibility of the feature distribution, the rationality of details, and the distortion situation of the area. This precise scanning process ensures the comprehensive identification of the feature information. Subsequently, the system further subdivides and reviews these features, and finally forms a comprehensive credibility analysis result.

[0147] According to the analysis results, check the feature information marked in the image, conduct a review based on pre-trained data or a preset artificial intelligence model, delete the images that do not meet the standards, and obtain the review confirmation results can be achieved through the following process:

[0148] After obtaining the preliminary credibility analysis results, the review system starts to conduct a comprehensive inspection of the enhanced image information. In this process, the system compares each marked feature information in the image with the feature standards issued within the organization. Through detailed comparison and verification work, it ensures that the feature information in the image meets the specified standards and the image enhancement effect is real and credible. This series of precise comparison and review operations enables the identification and elimination of non-compliant images, thereby ensuring the purity of the image content. Through this meticulous inspection process, the accuracy and reliability of the review confirmation results of the image are finally ensured.

[0149] The specific process of inserting a digital watermark into the image, setting up a monitoring program to track the distribution and use of the image, ensuring that the source and distribution records of the image are queryable, and obtaining anti-counterfeiting image data based on the review confirmation results is as follows:

[0150] According to the results confirmed by the review, immediately start embedding digital watermarks in the enhanced images. This operation is completed through a specially designed encoding technique, ensuring that both the pattern and the embedding location of the digital watermark are precisely calculated, guaranteeing the security of the image while ensuring that the image quality is not affected. At the same time, the system sets up an image monitoring program to monitor the usage of images in real time. This program can record the access and usage details of each image, ensuring that the source and purpose of the image can be traced and verified. Through this series of operations, a complete image enhancement and anti-counterfeiting data system is formed.

[0151] Embodiment 2

[0152] Based on the same inventive concept, this embodiment discloses a low-light image enhancement and security device based on deep learning. Please refer to Figure 2 , including:

[0153] A data collection and preprocessing module 201, which is used to collect a large number of low-light images covering different scenarios, lighting conditions, and object contents, and preprocess the collected low-light images;

[0154] A model construction module 202, which is used to construct a deep learning model. The deep learning model includes an image light simulation module and an image enhancement module. Among them, the image light simulation module is used to simulate the influence of the light source on the image based on the low-light image data, calculate the scattering, reflection, and shadow effects after the interaction between the light and the image, and generate light simulation data. The image enhancement module is used to adjust relevant parameters for enhancement processing according to the light simulation data and output the enhanced image. The image enhancement module adopts a combination of mean square error and structural similarity loss functions to guide the model to pay attention to the image structure and detail preservation while increasing the brightness, and introduces a perceptual loss function to ensure that the enhanced image is similar to the reference image in terms of high-level semantic features;

[0155] A training module 203, which is used to train the constructed deep learning model with the preprocessed low-light image data;

[0156] An image processing module 204, which is used to enhance the low-light image to be processed by using the trained deep learning model.

[0157] It should be noted that the image processing module 204 is the specific application or test of the model, including two modules: image light simulation and image enhancement. The specific functions of the two modules are the same as those of the two modules in the model construction module 202.

[0158] In one embodiment, the device further includes an artificial intelligence security and image credibility module 205, which is used to perform content review on the enhanced image, evaluate its authenticity and credibility by analyzing the visibility of the feature distribution, the rationality of details, and the distortion of the enhanced image, and at the same time generate image data with anti-counterfeiting characteristics by embedding digital watermarks and performing image content monitoring.

[0159] In one embodiment, please refer to Figure 5 , the artificial intelligence security and image credibility module 205 includes:

[0160] A credibility analysis sub-module 2051, which is used to screen the feature information of the image based on the enhanced image, identify the visibility of the feature distribution, the rationality of details, and the distortion, obtain an analysis result, and measure the true credibility of the enhanced image.

[0161] A content review sub-module 2052, which is used to check the identified feature information in the image according to the analysis result, perform review according to pre-trained data or a preset artificial intelligence model, delete the images that do not meet the standards, and obtain a review confirmation result.

[0162] An anti-counterfeiting coding sub-module 2053, which is used to insert a digital watermark into the image according to the review confirmation result, set up a monitoring program to track the distribution and use of the image, ensure that the source and distribution records of the image are queryable, and obtain anti-counterfeiting image data.

[0163] Since the device introduced in the second embodiment of the present invention is the device used to implement the deep learning-based low-light image enhancement security method in the first embodiment of the present invention, based on the method introduced in the first embodiment of the present invention, those skilled in the art can understand the specific structure and variations of the device, so it will not be elaborated here. Any device used in the method of the first embodiment of the present invention belongs to the scope protected by the present invention.

[0164] Embodiment Three

[0165] Based on the same inventive concept, please refer to Figure 3 , the present invention also provides a computer-readable storage medium 300, on which a computer program 311 is stored, and when the program is executed by a processor, it implements the method described in Embodiment One.

[0166] Since the computer-readable storage medium introduced in the third embodiment of the present invention is the computer-readable storage medium used to implement the deep learning-based low-light image enhancement security method in the first embodiment of the present invention, based on the method introduced in the first embodiment of the present invention, those skilled in the art can understand the specific structure and variations of the computer-readable storage medium, so it will not be elaborated here. Any computer-readable storage medium used in the method of the first embodiment of the present invention belongs to the scope protected by the present invention.

[0167] Embodiment 4

[0168] Based on the same inventive concept, please refer to Figure 4 , the present invention also provides a computer device, including a memory 401, a processor 402, and a computer program 403 stored on the memory and executable on the processor. When the processor executes the program, the method described in Embodiment 1 is implemented.

[0169] Since the computer device introduced in Embodiment 4 of the present invention is the computer device adopted for implementing the low-light image enhancement security method based on deep learning in Embodiment 1 of the present invention, based on the method described in Embodiment 1 of the present invention, those skilled in the art can understand the specific structure and variations of this computer device, so it will not be elaborated here. Any computer device adopted by the method in Embodiment 1 of the present invention falls within the scope of protection of the present invention.

[0170] Those skilled in the art should understand that the embodiments of the present invention can be provided as a method, a system, or a computer program product. Therefore, the present invention can be implemented in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention can be implemented in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk memories, CD-ROMs, optical memories, etc.) containing computer-usable program codes.

[0171] The present invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each flow and / or block in the flowchart and / or block diagram can be implemented by computer program instructions, and the combination of the flows and / or blocks in the flowchart and / or block diagram can also be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the specified functions in Figure 1 one or more of the flows and / or multiple flows and / or blocks Figure 1 one or more of the blocks and / or multiple blocks.

[0172] Although the preferred embodiments of the present invention have been described, additional changes and modifications can be made to these embodiments by those skilled in the art once they learn the basic creative concept. Therefore, the appended claims are intended to be construed to include the preferred embodiments as well as all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art can make various changes and variations to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Thus, if these modifications and variations of the embodiments of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these changes and variations.

Claims

1. A low-light image enhancement security method based on deep learning, characterized in that: include: Collect a large number of low-light images covering different scenes, lighting conditions and object contents, and pre-process the collected low-light images; A deep learning model is constructed, and the deep learning model includes an image light simulation module and an image enhancement module, wherein the image light simulation module is used to simulate the influence of the light source on the image based on the low-brightness image data, calculate the scattering, reflection and shadow effects after the interaction between the light and the image, and generate light simulation data; the image enhancement module is used to adjust the relevant parameters for enhancement processing according to the light simulation data, and output the enhanced image. The model adopts a combination of mean square error and structural similarity loss function to guide the model to focus on image structure and detail retention while improving brightness, and introduces a perceptual loss function to ensure that the enhanced image is similar to the reference image in high-level semantic features. The image enhancement module includes a parameter adjustment submodule, a detail enhancement submodule and an image output submodule, wherein the parameter adjustment submodule is used to analyze the color distribution and light intensity in the image based on the light simulation data, adjust the parameters one by one to match the predetermined visual effect, and adjust the brightness level and color saturation to generate image parameter optimization data, the detail enhancement submodule is used to use the image parameter optimization data to perform hierarchical separation, sharpness and contrast adjustment on the details in the image, and optimize the key visual elements in the image, and the image output submodule is used to adjust the resolution and perform color correction through the detail optimized image, and output the enhanced image; The constructed deep learning model is trained using the preprocessed low-brightness image data; Use the trained deep learning model to enhance the low-light images to be processed; The enhanced image is audited for content, and its authenticity and credibility are evaluated by analyzing the feature distribution visibility, detail rationality, and distortion of the enhanced image. At the same time, image data with anti-counterfeiting properties is generated by embedding digital watermarks and performing image content monitoring.

2. The low-light image enhancement security method based on deep learning as claimed in claim 1, characterized in that: Preprocessing of the original low-light image data includes: Coordinate calibration, resolution adjustment and format conversion are performed on the original low-light image to generate standardized pre-processed low-light image data.

3. The low-light image enhancement security method based on deep learning as claimed in claim 1, characterized in that: The preprocessed low-light image data is used to train the constructed deep learning model, including: Construct a training dataset based on the preprocessed low-brightness image data; The training data set is input into the constructed deep learning model, and the model parameters are optimized by calculating the difference between the model prediction and the actual expectation. In each iteration, the gradient is calculated according to the loss function, and then the weights and bias of the network are updated using the back propagation algorithm. The loss function is constructed based on the mean square error, structural similarity loss function, and perceptual loss function.

4. The low-light image enhancement security method based on deep learning as claimed in claim 1, characterized in that: The model uses a combination of mean square error and structural similarity loss functions to guide the model to focus on image structure and detail preservation while improving brightness, including: The mean square error loss function is used to measure the difference between the enhanced image and the reference normal illumination image: in, is the total number of image pixels, is the i-th pixel of the input low-light image, is the i-th pixel of the enhanced image output by the model, is the mean square error loss function; The structural similarity loss function is introduced to measure the structural similarity between two images: in, is the input low-light image, is the enhanced image output by the model, It is a function to measure the structural similarity between two images. The closer the value is to 1, the more similar the image structures are, the smaller the image distortion is, and the more similar and reliable the enhancement effect is. is the structural similarity loss function; The perceptual loss function is: in, is the number of feature layers, Respectively l The number of channels, height and width of the layer feature map, is the reference image, is the enhanced image output by type, It is l The perceptual features of the enhanced image, It is l The perceptual features of the reference images, Indicates the number of channels, height, and width of the image.

5. The low-light image enhancement security method based on deep learning as claimed in claim 1, characterized in that: The detail enhancement submodule is specifically used for: Use the gradient information of the image to construct a detail feature map; Based on the detail feature map, a multi-scale processing method is used to enhance the details. For each scale of the detail feature map, a detail enhancement filter is designed for processing. A noise suppression mechanism is introduced in the detail enhancement process. The degree of detail enhancement is adaptively adjusted according to the local variance or noise estimation value of the image. The feature maps of different scales after detail enhancement are fused to obtain the detail enhancement result; The detail enhancement result is fused with the image after brightness and color adjustment to obtain the final enhanced image.

6. A low-light image enhancement safety device based on deep learning, characterized in that: include: The data collection and preprocessing module is used to collect a large number of low-light images covering different scenes, lighting conditions and object contents, and preprocess the collected low-light images; The model building module is used to build a deep learning model, and the deep learning model includes an image light simulation module and an image enhancement module. The image light simulation module is used to simulate the influence of the light source on the image based on the low-light image data, calculate the scattering, reflection and shadow effects after the interaction between the light and the image, and generate light simulation data. The image enhancement module is used to adjust the relevant parameters for enhancement processing according to the light simulation data, and output the enhanced image. The image enhancement module adopts a combination of mean square error and structural similarity loss function to guide the model to focus on image structure and detail retention while improving brightness, and introduces a perceptual loss function to ensure that the enhanced image is in high-level semantic features. Similar to the reference image, the image enhancement module includes a parameter adjustment submodule, a detail enhancement submodule and an image output submodule, wherein the parameter adjustment submodule is used to analyze the color distribution and light intensity in the image based on the light simulation data, adjust the parameters one by one to match the predetermined visual effect, and adjust the brightness level and color saturation to generate image parameter optimization data, the detail enhancement submodule is used to use the image parameter optimization data to perform hierarchical separation of details in the image, adjust the sharpness and contrast, and optimize the key visual elements in the image, and the image output submodule is used to adjust the resolution and perform color correction through the detail optimized image, and output the enhanced image; A training module is used to train the constructed deep learning model using the preprocessed low-brightness image data; Image processing module, used to enhance the low-brightness image to be processed using the trained deep learning model; The artificial intelligence security and image trustworthiness module is used to perform content review on the enhanced images. It evaluates the authenticity and credibility of the enhanced images by analyzing their feature distribution visibility, detail rationality, and distortion. It also generates image data with anti-counterfeiting properties by embedding digital watermarks and performing image content monitoring.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the low-light image enhancement security method based on deep learning as described in any one of claims 1 to 5 is implemented.

8. A computer 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 program, the low-light image enhancement security method based on deep learning as described in any one of claims 1 to 5 is implemented.

Images