Physically-guided low-light image enhancement method based on joint modeling of illumination exposure

Abstract

The present application relates to the technical field of image processing, and discloses a physical guiding low-light image enhancement method based on joint modeling of illumination and exposure, wherein a low-light image to be processed is input into an illumination and exposure prediction module to perform feature extraction and generate a local illumination map, and a global exposure factor is derived based on the local illumination map; the low-light image to be processed, the local illumination map and the global exposure factor are input into a double condition adaptive enhancement module, intensity modulation is performed in combination with the global exposure factor, spatially guided interactive operation is performed in combination with the local illumination map, and an enhanced image is output after feature reconstruction; in the model training stage, a normal light reference image is acquired, reconstruction loss and structural consistency loss are calculated, exposure consistency loss is calculated for the global exposure factor, and an overall loss function is constructed to drive an optimizer to jointly optimize network parameters. The present application realizes adaptive calibration of local space and global brightness, effectively suppresses image noise while recovering dark details.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a physically guided low-light image enhancement method based on joint modeling of illumination and exposure. Background Technology

[0002] In practical applications, due to insufficient ambient light or equipment exposure limitations, captured images often suffer from low overall brightness, poor contrast, and loss of detail in dark areas. Low-light image enhancement techniques aim to restore the visibility of these degraded images. Current deep learning image enhancement methods have limitations in handling the relationship between local spatial illumination and global brightness. Many existing techniques rely on a single parameter for overall brightness adjustment or treat local illumination and global exposure as two independent branches, severing the physical connection between them. This approach cannot accommodate the brightness range of different areas within the same image. Forcibly increasing the brightness of dark areas can lead to overexposure in areas that were originally well-lit; conversely, attempting to suppress overexposure can result in insufficient brightening of dark areas, making it difficult to achieve uniform brightness restoration across the entire image.

[0003] In the feature extraction and reconstruction stages, existing networks typically employ an indiscriminate feature processing strategy for the entire image. Due to the lack of spatial prior guidance on local illumination distribution and the absence of effective feature intensity modulation mechanisms for both global and local aspects, the network cannot allocate corresponding processing weights based on the actual degradation degree of different regions. Consequently, when existing algorithms attempt to recover texture details in extremely dark areas, they simultaneously amplify the hidden background noise in those areas. The processed images are prone to color distortion and structural blurring, making it difficult to maintain the overall color and structural consistency of the image.

[0004] Furthermore, existing model training criteria largely rely on conventional pixel-level absolute differences or structural similarity to construct loss functions, leaving the evaluation of image exposure deviations at the level of direct linear calculation. Because the human visual system's perception of brightness is non-linear, existing linear evaluation methods fail to incorporate the physiological perception patterns of the human eye into the network constraints. This results in the model lacking accurate physical exposure correction criteria during training, ultimately producing enhanced images that appear stiff and unnatural in their visual brightness distribution, failing to truly conform to the real physiological perception habits of the human eye. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a physically guided low-light image enhancement method based on joint modeling of illumination and exposure. This method solves the problems in existing low-light image enhancement techniques, such as overexposure in local areas or insufficient brightening of dark areas due to the severing of the physical relationship between local illumination and global exposure, and amplification of noise and color distortion when restoring details in dark areas due to the lack of spatial prior guidance.

[0006] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a physically guided low-light image enhancement method based on joint modeling of illumination and exposure, comprising the following steps: The low-light image to be processed is input into the illumination and exposure prediction module, and the low-light image to be processed is subjected to feature extraction to generate a local illumination map. The global exposure factor is then derived serially from the local illumination map. The low-light image to be processed, the local illumination map, and the global exposure factor are input into the dual-condition adaptive enhancement module. The global exposure factor is combined with intensity modulation, and the local illumination map features are combined with spatial guided interactive operation. After feature reconstruction, the enhanced image is output. During the model training phase, a normal light reference image is acquired. Reconstruction loss and structural consistency loss are calculated for the output enhanced image and the normal light reference image. Exposure consistency loss is calculated for the derived global exposure factor. An overall loss function is constructed, and the optimizer is driven to jointly optimize the network parameters based on the overall loss function.

[0007] Preferably, the processing method for extracting features from the low-light image to generate a local illumination map includes: capturing multi-scale features from the input low-light image using a backbone architecture, and using the multi-scale features as input features in the current processing stage; transmitting the input features to the illumination prediction block, performing layer normalization on the input features using the illumination prediction block, generating query features, key features, and value features respectively through independent linear transformation operations on the features that have completed layer normalization, and performing self-attention calculation using the query features, key features, and value features; inputting the features that have completed self-attention calculation into a feedforward network for nonlinear mapping processing to obtain the output features of the illumination prediction block; stacking the illumination prediction block into four layers, using convolutional layers to map the stacked output features to generate multi-channel features, compressing the multi-channel features in the channel dimension to obtain single-channel features, and outputting the single-channel features as a local illumination map.

[0008] In this step, multi-layered, continuously stacked illumination prediction blocks, combined with self-attention computation, can aggregate illumination information at different scales within the global receptive field. Subsequent channel compression forces the high-dimensional abstract features down to a single channel, effectively removing texture and color interference and accurately outputting a spatial illumination distribution matrix that conforms to physical meaning.

[0009] Preferably, the processing method for serially deriving the global exposure factor from local illumination map features includes: inputting the local illumination map features output from the network's end into the exposure derivation block; using multiple convolutional layers within the exposure derivation block to downsample the spatial dimension and compress the channel dimension features of the local illumination map features; using a linear rectified function to perform nonlinear mapping on the features output by convolution between adjacent convolutional layers; using pooling operations to compress the spatial dimension of the feature map and extract feature vectors representing global information; mapping the feature vectors into a single scalar value and inputting the scalar value into an activation function to restrict the value to a preset range; and outputting the global exposure factor in real number form after the activation function mapping process. This serial derivation mechanism enables the network to perceive the overall exposure deviation from the locally spatially varying illumination distribution, ensuring that the exposure adjustment parameters are physically reasonable.

[0010] Preferably, the intensity modulation processing method combined with the global exposure factor includes: using a multilayer perceptron built inside the dual-condition adaptive enhancement module to perform nonlinear mapping on the global exposure factor to generate a first set of affine transformation parameters and a second set of affine transformation parameters; extracting features from the low-light image to be processed to obtain low-light input features; extracting the first set of affine transformation parameters to perform a feature dimension multiplication operation on the low-light input features to adjust the response amplitude of the features; and simultaneously using another parameter in the first set of affine transformation parameters to perform an addition operation on the low-light input features after the multiplication operation to provide an offset of the feature reference value, thereby generating low-light modulation features.

[0011] Preferably, the processing method for performing spatially guided interactive operations in conjunction with local illumination map features includes: performing layer normalization on low-light modulation features; using independent linear transformation layers to map features to feature subspaces and combining with dimension reshaping operations to generate query features, key features, and value features respectively; calculating the product of the transpose matrix of the key feature and the matrix of the query feature and introducing a scaling factor to adjust the numerical value of the product result; using a normalized exponential function to map the adjusted result into normalized spatially guided weights; using the local illumination map as the local illumination map feature input to the convolutional layer for feature preprocessing; applying the spatially guided weights to the product of the convolutionally processed local illumination map features and the value features; and adding the interactive features to the low-light modulation features through residual connections to output the spatially guided features.

[0012] The above interactive operation uses the local illumination map as an explicit spatial prior condition to guide the network to allocate different self-attention recovery weights in different brightness regions, enabling the algorithm to adaptively apply the corresponding feature repair strategies in extremely dark areas and normal bright areas.

[0013] Preferably, the processing method for the enhanced image output after feature reconstruction includes: performing layer normalization on the spatial guiding features; inputting the normalized spatial guiding features into a feedforward network for nonlinear feature reshaping; adding the output of the feedforward network with the original spatial guiding features by residual addition to generate refined features; extracting the second set of affine transformation parameters generated by the multilayer perceptron mapping, which includes a reconstruction scaling factor and a reconstruction offset factor; using the reconstruction scaling factor to perform feature dimension multiplication on the refined features to regulate the feature response intensity in the reconstruction stage; simultaneously using the reconstruction offset factor to perform addition on the refined features after multiplication to achieve secondary calibration of the feature baseline value and generate reconstruction intensity features; transmitting the generated reconstruction intensity features to the projection layer at the output end; using a convolutional layer to project the multidimensional feature tensor back to the image color space in the channel dimension; connecting an activation function after the output of the projection layer to limit the output value to the normalized range of image pixel values, and finally outputting the enhanced image.

[0014] Preferably, the processing method for calculating exposure consistency loss based on the derived global exposure factor includes: calculating the average exposure value of the low-light image and the normal-light reference image in the global dimension, summing all pixel values ​​in the color channel dimension and the two-dimensional space dimension, and averaging them with the total number of pixels to obtain the global exposure average; calculating the actual exposure difference based on the extracted global exposure average, transforming the global exposure average to logarithmic space for difference calculation, and introducing a smoothing constant into the input term of the logarithmic function; comparing the derived global exposure factor with the actual exposure difference calculated in logarithmic space, and using the norm to calculate the absolute difference to construct the exposure consistency loss. The numerical distribution pattern in logarithmic space highly matches the brightness perception characteristics of the human visual system. Difference calculation based on logarithmic space can directly guide the network to output a physical exposure deviation correction value that closely resembles human visual habits.

[0015] Preferably, the processing method for calculating reconstruction loss and structural consistency loss between the output enhanced image and the normal light reference image and constructing an overall loss function includes: calculating the reconstruction loss by using the norm to calculate the absolute difference between the pixels at corresponding spatial locations in the enhanced image and the normal light reference image; inputting the enhanced image and the normal light reference image into a structural similarity function to calculate the similarity value and converting it into structural consistency loss; and linearly combining the exposure consistency loss, reconstruction loss and structural consistency loss, and setting a corresponding weight coefficient for each loss to construct an overall loss function.

[0016] Preferably, the processing method of jointly optimizing network parameters based on the overall loss function-driven optimizer includes: using an adaptive moment estimation optimizer to jointly optimize network parameters using the overall loss function; setting an initial learning rate and a corresponding dynamic learning rate decay strategy, and introducing a cosine annealing strategy to dynamically adjust the initial learning rate so that the learning rate exhibits a non-linear decreasing trend as the training cycle increases and smoothly decays to the lower limit value at the end of the set maximum training cycle.

[0017] This invention provides a physically guided low-light image enhancement method based on joint modeling of illumination and exposure. It has the following beneficial effects: 1. This invention extracts features from the low-light image to be processed through an illumination and exposure prediction module. First, a local illumination map is generated, and then the global exposure factor is derived serially based on the local illumination map. This combines the illumination distribution in the local spatial dimension with the exposure compensation in the global dimension. The serial derivation mechanism ensures that the global exposure factor is based on the physical foundation of the local illumination features, thereby solving the problem of overexposure in local areas or insufficient brightening of dark areas that can easily be caused by adjusting a single parameter, and realizing the uniform restoration of the brightness of the entire image.

[0018] 2. This invention constructs a dual-condition adaptive enhancement module, which uses the global exposure factor for affine transformation modulation of feature intensity, and uses the local illumination map as a spatial prior to participate in self-attention calculation. In the feature reconstruction stage, the network uses global parameters to adjust the response amplitude and reference offset of features, and uses local spatial weights to allocate recovery weights for regions with different brightness, so that the network can specifically process regions with different degrees of degradation. While restoring the texture details of dark areas, it suppresses noise amplification and maintains the color and structural consistency of the processed image.

[0019] 3. In the model training phase, this invention constructs an overall loss function that includes exposure consistency loss. The global exposure mean of the image is transformed into logarithmic space to calculate the true exposure difference, and compared with the global exposure factor derived by the network. The logarithmic calculation mode conforms to the nonlinear perception characteristics of human visual system for brightness. By using this loss function to constrain the network, physical exposure deviation can be corrected. Combined with reconstruction loss and structural consistency loss, network parameters are jointly optimized so that the enhanced image output by the network conforms to the real physiological perception habits of human eyes in terms of visual brightness distribution. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall architecture of the physically guided low-light image enhancement system based on joint modeling of illumination and exposure according to the present invention. Figure 2 This is an overall flowchart of the physical-guided low-light image enhancement method based on joint modeling of illumination and exposure, as proposed in this invention. Detailed Implementation

[0021] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See attached document Figure 1 This invention provides a physically guided low-light image enhancement method based on joint modeling of illumination and exposure, which includes the following system structure and process steps.

[0023] The low-light image enhancement system provided by this invention receives low-light images. As input data, output the corresponding enhanced image. .in, This indicates that the low-light image belongs to the real number space. The number of channels is Height is And the width is The tensor. This indicates that the enhanced image belongs to the real space. The number of channels is Height is And the width is The tensor of the low-light image enhancement system includes an illumination and exposure prediction module and a dual-condition adaptive enhancement module.

[0024] The illumination and exposure prediction module is used to predict the light from the input low-light image. Extract brightness distribution features to generate a local illumination map. .in, This indicates that the local illumination map belongs to the real space. Single channel, height is And the width is The tensor of the illumination exposure prediction module further derives the global exposure factor from the local illumination map based on a serial derivation mechanism. ,in, , indicating that the global exposure factor belongs to the real number space.

[0025] The dual-condition adaptive enhancement module is connected to the illumination exposure prediction module. It receives the global exposure factor and local illumination map output by the illumination exposure prediction module, and uses the received global exposure factor and local illumination map to modulate and guide the image features of the low-light image, reconstruct and output the enhanced image.

[0026] Based on the above system structure, the overall workflow of the physically guided low-light image enhancement method based on joint modeling of illumination and exposure includes the following steps.

[0027] Step S1: Obtain the low-light image to be processed and input it into the illumination and exposure prediction module. The illumination and exposure prediction module extracts features from the low-light image to generate a local illumination map. Then, through the convolutional layer, pooling operation and activation function of the internal exposure derivation block, the global exposure factor is serially derived from the features of the local illumination map.

[0028] In step S2, the low-light image, local illumination map, and global exposure factor are simultaneously input into the dual-condition adaptive enhancement module. On the one hand, the dual-condition adaptive enhancement module uses the global exposure factor to generate affine transformation parameters through multilayer perceptron mapping. At the network front end and back end, primary intensity modulation and global intensity modulation are performed on the features extracted from the low-light image, respectively. On the other hand, the cross-attention between the local illumination map features and the low-light modulation features is calculated to perform spatially guided interactive operations. Then, the refined features after residual reshaping and the affine transformation parameters are combined to perform feature reconstruction. Finally, the enhanced image is projected and output.

[0029] Step S3: During the model training phase, an overall loss function is constructed. Specifically, reconstruction loss and structural consistency loss are calculated for the output enhanced image and the corresponding normal light reference image. At the same time, for the global exposure factor output by the illumination exposure prediction module, the exposure consistency loss in the logarithmic domain is calculated by combining the global exposure mean of the normal light reference image and the low light image. Finally, based on the overall loss function obtained by weighted summation of various losses, the optimizer is driven to jointly optimize the network parameters of the illumination exposure prediction module and the dual-condition adaptive enhancement module.

[0030] See attached document Figure 2 The illumination prediction module extracts features from low-light images to generate local illumination maps. The specific operation can be divided into the following steps.

[0031] Step S101: The illumination exposure prediction module adopts a backbone architecture based on the U-Net network, utilizing this encoding and decoding structure to predict the input low-light image. Multi-scale features are captured. These multi-scale features serve as input features for the current processing stage. Regarding the specific downsampling, upsampling, and skip connection operations of the U-Net network, those skilled in the art can make conventional settings according to actual feature extraction requirements. Its basic network connection form is a well-known technology in the field and will not be described in detail here.

[0032] Step S102: The illumination prediction module integrates an illumination prediction block at each layer of the network architecture to predict illumination features at different resolutions. The illumination prediction module extracts the input features... The data is then transmitted to the corresponding illumination prediction block. The illumination prediction block processes the input features. A layer normalization operation is performed to standardize the feature distribution. The features after layer normalization are then transformed through independent linear transformations to generate query features for self-attention computation. Key features Sum value characteristics The illumination prediction block utilizes query features. Key features and value characteristics Perform self-attention calculations to capture illumination information at different scales.

[0033] In step S103, the illumination prediction block inputs the features that have completed self-attention calculation into the internal feedforward network. The feedforward network performs nonlinear mapping on the input features, ultimately obtaining the output features of this illumination prediction block. .

[0034] In step S104, the illumination prediction module stacks the illumination prediction blocks of the above structure into four layers. The illumination prediction module controls the features to flow and process sequentially through the four consecutive illumination prediction blocks to extract deep illumination information. After processing by the four stacked illumination prediction blocks, the illumination prediction module maps the features using a convolutional layer of a specific size. This convolutional layer uses a 3×3 kernel to compress multi-channel features along the channel dimension and output them as single-channel features, generating a local illumination map. The generated local illumination map Used to characterize low-light images The distribution of regional brightness in two-dimensional space provides a physical guide for subsequent modules at the local spatial level.

[0035] The illumination exposure prediction module derives the global exposure factor based on a serial derivation mechanism. The specific operation can be divided into the following steps.

[0036] In step S105, after obtaining the extracted features of the local illumination map, the illumination exposure prediction module generates the global exposure factor using a serial derivation mechanism. The extracted features of the local illumination map are used to characterize the regional brightness distribution in the spatial dimension, and the global exposure factor. This is used to quantify the macroscopic exposure level of the entire low-light image. Parallel extraction of local spatial features and global scalar features can easily lead to feature conflicts during the network learning process. The illumination and exposure prediction module directly derives the global exposure factor based on the local illumination map features extracted in the aforementioned steps. Establish a physical correspondence between local brightness distribution and global exposure level to ensure that the two extracted physical quantities have inherent consistency.

[0037] In step S106, the illumination exposure prediction module inputs the local illumination map features output from the network's end into the internal exposure derivation block. The exposure derivation block consists of multiple convolutional layers cascaded with activation functions. The exposure derivation block utilizes these convolutional layers to downsample the local illumination map features in both spatial and channel dimensions. Between adjacent convolutional layers, the exposure derivation block uses a linear rectified function as the activation function to perform a non-linear mapping on the features output by the convolution. The specific kernel size, number of network layers, and stride parameters for the multiple convolutional layers can be conventionally set by those skilled in the art based on actual feature dimensionality reduction requirements. The configuration of the basic network layers is well-known in the field and will not be elaborated upon here.

[0038] In step S107, after completing multi-layer convolution and non-linear mapping, the exposure derivation block uses pooling operations to compress the spatial dimension of the feature map and extract a feature vector representing global information. The exposure derivation block maps this feature vector to a single scalar value and inputs this scalar value into the Sigmoid activation function. The Sigmoid activation function restricts the range of the input value to a preset range of consecutive real numbers from 0 to 1. After the Sigmoid activation function mapping process, the exposure derivation block outputs a global exposure factor in real-valued form. Global exposure factor The image is then transmitted to the dual-condition adaptive enhancement module as a physical guide parameter for adjusting the overall exposure intensity of the input image.

[0039] See attached document Figure 2 The dual-condition adaptive enhancement module performs dynamic feature modulation on image features based on the global exposure factor. The specific operation can be divided into the following steps.

[0040] Step S201: The dual-condition adaptive enhancement module receives the global exposure factor output by the illumination exposure prediction module. Then, the global exposure factor is analyzed using an internally built multilayer perceptron. A nonlinear mapping is performed to generate a set of affine transformation parameters. The multilayer perceptron transforms the single-scalar-dimensional global exposure features into multi-dimensional network modulation weights. The specific mapping calculation process is shown in the following equation: ; in, Represents the global exposure factor; Represents a multilayer perceptron; , , , These represent four independent affine transformation parameters generated by the multilayer perceptron mapping. The specific number of fully connected layers, the number of hidden layer neurons, and the internal activation functions in the multilayer perceptron can be conventionally set by those skilled in the art according to the actual parameter mapping dimension requirements. The basic network construction method is a well-known technique in this field and will not be elaborated upon here.

[0041] In step S202, the dual-condition adaptive enhancement module divides the generated affine transformation parameters into two groups based on the different stages of network processing. The first group of affine transformation parameters includes... and This is configured at the front end of the dual-condition adaptive enhancement module and is used to perform primary intensity modulation on the input low-light image features. The second set of affine transformation parameters includes... and This is configured at the end of the dual-condition adaptive enhancement module, used for final intensity modulation during the feature reconstruction stage after spatial guidance and feature interaction are completed. This is achieved by using a global exposure factor that characterizes the macroscopic exposure level. By converting the parameters into affine transformation parameters for the network layer, the system achieves dynamic control of the distribution of local features based on the overall illumination conditions of the image.

[0042] Step S203: The dual-condition adaptive enhancement module acquires the low-light input features extracted from the low-light image. The dual-condition adaptive enhancement module extracts the first set of affine transformation parameters and uses these parameters... Low light input characteristics Multiplication operations are performed along the feature dimensions to adjust the response amplitude of the features; simultaneously, affine transformation parameters are utilized. An addition operation is performed on the features after the multiplication operation to provide an offset for the feature reference value. The dual-condition adaptive enhancement module completes the primary intensity modulation through the above operation to generate low-light modulation features. The specific calculation process for primary intensity modulation is shown in the following formula: ; in, This indicates the low-light modulation characteristics generated after primary intensity modulation; This represents the low-light input features extracted from the low-light image. The dual-condition adaptive enhancement module performs preliminary calibration of the brightness distribution range of the low-light image features at the feature space level through this affine transformation process.

[0043] The dual-condition adaptive enhancement module performs spatially guided interaction on image features based on local illumination maps. The specific operation can be divided into the following steps.

[0044] In step S204, after completing the primary intensity modulation, the dual-condition adaptive enhancement module receives the local illumination map extracted by the illumination exposure prediction module. and use it as a feature of the illumination map. It participates in subsequent calculations. The dual-condition adaptive enhancement module utilizes illumination map features. With low light modulation characteristics Perform guided interaction in the spatial dimension. The spatial guided interaction process achieves targeted reconstruction of local image details by establishing a correlation between local brightness distribution and low-light image features.

[0045] Step S205, the dual-condition adaptive enhancement module first applies low-light modulation features The execution layer performs a normalization operation to stabilize the feature distribution. Then, independent linear transformation layers are used to map the features to specific feature subspaces, and dimensionality reshaping operations are used to generate query features. Key features Sum value characteristics The specific generation process is shown in the following formula: ; ; ; in, Presentation layer normalization operation; Indicates a linear transformation layer; This indicates a reshaping operation of the feature dimensions; Indicates query characteristics; Indicates key features; This represents the value feature. The dual-condition adaptive enhancement module uses this operation to transform the low-light modulation feature. This can be transformed into different representations that can be used to compute self-attention.

[0046] Step S206: The dual-condition adaptive enhancement module performs spatial guidance using a cross-attention mechanism. The dual-condition adaptive enhancement module calculates key features. Transpose matrix and query features The matrix product is used to characterize the spatial correlation within the image and introduces a scaling factor. Adjust the product result numerically. Scaling factor. The value is based on query features Key features Feature Dimensions Confirmed. Specifically, Indicate query characteristics Bond features Dimension size in the channel direction, scaling factor The value can be taken as the feature dimension. The reciprocal of the square root is used to prevent gradient vanishing during backpropagation of attention weights. The dual-condition adaptive enhancement module utilizes... The function maps the adjusted result to normalized spatial guided weights.

[0047] Step S207, the dual-condition adaptive enhancement module will enhance the illumination map features Input a specific convolutional layer for feature preprocessing, making it correlated with the value features in the channel dimension. Matching. The dual-condition adaptive enhancement module applies spatially guided weights to the convolutionally processed illumination map features and value features. The product of these features is then used to inject local illumination distribution information into the feature interaction process. Finally, the interacted features are combined with the low-light modulation features via residual connections. Addition, output spatial guided features The specific spatially guided interactive calculation process is shown in the following formula: ; in, Indicates spatial guidance characteristics; Indicates the characteristics of the illumination map; Indicates a convolutional layer. This represents the scaling factor used to adjust the attention weights; This represents the normalized exponential function; Key features The transpose of the matrix. The dual-condition adaptive enhancement module, through the aforementioned spatially guided interaction, ensures that the enhancement process can adaptively repair features and restore details based on differences in local illumination intensity.

[0048] After completing the spatial guided interaction, the dual-condition adaptive enhancement module performs feature reconstruction and generates the final enhanced image. The specific operation can be divided into the following steps.

[0049] Step S208: The dual-condition adaptive enhancement module acquires the spatial guidance features generated during the spatial guidance interaction process. Dual-condition adaptive enhancement module for spatial guidance features A layer normalization operation is performed, and then the normalized features are input into a feedforward network for nonlinear feature reshaping. The feedforward network achieves depth mapping of features through two linear transformation layers and an intermediate activation function layer to extract fine details and color information from the image. The dual-conditional adaptive enhancement module combines the output of the feedforward network with the original spatially guided features. Perform residual summation to generate refined features. The specific calculation process for the refining characteristics is shown in the following formula: ; in, This indicates the generated refined features; Indicates a feedforward network; Presentation layer normalization operation; This represents the spatially guided features. The dual-condition adaptive enhancement module, through this residual structure, further enhances the expressive power of the features while preserving the physical constraints of spatial guidance.

[0050] In step S209, the dual-condition adaptive enhancement module uses the second set of affine transformation parameters generated in step S201 to refine the features. Perform the final global intensity modulation. The dual-condition adaptive enhancement module extracts the values ​​obtained from the previous steps by the multilayer perceptron based on the global exposure factor. The second set of affine transformation parameters generated by the mapping and That is, the reconstruction scaling factor With reconstruction offset factor .

[0051] To ensure that the enhanced image brightness conforms to visual perception logic and does not cause pixel saturation overflow, the scaling factor is reconstructed. The value range is set to a real number interval greater than 0. In a specific embodiment, its value is distributed between [0.5, 5.0], and it is used to refine the features. Perform gain control; reconstruct offset factor The value range is set to a real number interval. In a specific embodiment, its value is distributed between [-1.0, 1.0], used to provide a linear shift of the brightness reference. The boundary of the above value range is determined by the network's adjustment of the overall loss function during the training phase. The result was determined by gradient descent optimization.

[0052] The dual-condition adaptive enhancement module utilizes the reconstruction scaling factor Refining characteristics Multiplication operations are performed along the feature dimensions to modulate the feature response intensity during the reconstruction stage; simultaneously, a reconstruction offset factor is utilized. An addition operation is performed on the features after the multiplication operation to achieve secondary calibration of the feature baseline values. The specific feature reconstruction calculation process is shown in the following formula: ; in, This represents the generated reconstruction intensity feature; Indicates a refined characteristic; Indicates the reconstruction scaling factor; This represents the reconstruction offset factor. The dual-condition adaptive enhancement module uses this step to adjust the global exposure factor. The underlying intention of macroscopic brightness control is ultimately realized through spatially guided enhancement features.

[0053] Step S210, the dual-condition adaptive enhancement module generates the reconstruction intensity features The data is transmitted to the projection layer at the output. This projection layer is a 3×3 convolutional layer that projects the multidimensional feature tensor back to the image color space along the channel dimension. The specific stride, padding, and weight initialization parameters of the convolutional layer can be conventionally set by those skilled in the art according to the resolution requirements of the output image. The basic convolutional projection operation is well-known in the field and will not be elaborated upon here.

[0054] In step S211, the dual-conditional adaptive enhancement module connects an activation function after the projection layer output to restrict the output values ​​to a normalized range of image pixel values, ultimately outputting the enhanced image. The activation function uses the Tanh function, which maps the feature responses to [ The pixel space of [1,1] is then normalized to restore it to the normal pixel value range of [0,255]. Image enhancement. It possesses color performance and texture details consistent with normal light reference images, achieving physical-guided low-light image quality improvement.

[0055] See attached document Figure 2 To ensure the model accurately learns the physical dimension of exposure mapping during training, the system constructs a multi-dimensional loss function to constrain the network. During model training, the system calculates the exposure consistency loss for the global exposure factor output by the illumination exposure prediction module. The specific operation can be divided into the following steps.

[0056] Step S301: The system acquires the input low-light image during the training phase. and the corresponding normal light reference image To provide accurate physical supervision signals for the network's predicted global exposure factor, the system calculates low-light images separately. and normal light reference image The global average exposure value is calculated by summing the pixel values ​​across all color channels and the two-dimensional space dimension, and then dividing by the total number of pixels. ; ; in, This represents the global average exposure value of the normal light reference image; This represents the global average exposure value of a low-light image. This represents a normal light reference image; Indicates a low-light image; This represents the operation of summing up all pixel values ​​in the color channel, height, and width dimensions of the normal light reference image; This represents the operation of summing up all pixel values ​​in the color channel, height, and width dimensions of a low-light image; These represent the pixel coordinate indices of the image tensor in the color channel, height, and width dimensions, respectively. Indicates the total number of color channels in an image; Indicates the height of the image; This represents the width of the image. The system uses this operation to extract a scalar value representing the macroscopic brightness level of the two images.

[0057] Step S302, the system based on the extracted global exposure average value and The system calculates the true exposure difference between the two values. To ensure the calculation results conform to the nonlinear perception characteristics of light intensity in the human visual system, the system transforms the exposure mean into logarithmic space for difference calculation. During the logarithmic transformation, to prevent the logarithmic function from becoming meaningless or overflowing due to zero input values, the system introduces a minimum smoothing constant into the input term of the logarithmic function.

[0058] Step S303: The system obtains the global exposure factor predicted by the illumination exposure prediction module during the forward propagation process. The system will use the predicted global exposure factor. The difference is compared with the actual exposure difference calculated in the logarithmic domain. The system uses the L1 norm to calculate the absolute difference between the two to construct the exposure consistency loss. The specific calculation process for exposure consistency loss is shown in the following formula: ; in, This indicates a loss of exposure consistency; This represents the global exposure factor predicted and output by the illumination prediction module; and These represent the global exposure averages of the normal light reference image and the low light image, respectively; This represents a smoothing constant used to prevent numerical computation overflow. Represents the natural logarithm function; This indicates the operation of calculating the L1 norm.

[0059] Step S304, for the smoothing constant in the above formula Its main function is to provide numerical stability. To avoid significant interference with the original exposure mean, a smoothing constant is used. The value of is set to a positive real number close to zero; in a specific embodiment, its fixed value is 10. -6 The system minimizes this exposure consistency loss. The forced illumination exposure prediction module learns the true exposure ratio mapping relationship between low-light and normal-light environments to ensure the predicted global exposure factor. It possesses reliable physical quantitative significance.

[0060] During the model training phase, in addition to constraining the exposure mapping relationship in the physical dimension, the system also needs to impose targeted constraints on image content at the pixel and perceptual levels in order to construct the overall loss function. The specific operations can be divided into the following steps.

[0061] Step S305: The system acquires the enhanced image output by the dual-condition adaptive enhancement module. and the normal light reference image in the corresponding training samples. To constrain the network at the global color and illumination levels, and to enable the network to generate images that approximate the real scene in pixel space, the system calculates image enhancement. Compared with normal light reference image The system calculates the reconstruction loss between the two images using the L1 norm to determine the absolute difference between pixels at corresponding spatial locations. The specific calculation process for the reconstruction loss is shown in the following formula: ; in, Indicates the losses incurred during reconstruction; This represents a normal light reference image; Indicates an enhanced image; This represents the operation of calculating the L1 norm. By minimizing this reconstruction loss, the system establishes a basic color alignment criterion for mapping low-light images to normal-light images.

[0062] Step S306: Calculating single pixel-level differences can easily lead to blurring at image edges and high-frequency details. The system introduces structural consistency loss on top of the reconstruction loss, evaluating the local similarity of the image from feature dimensions such as brightness, contrast, and structural distribution. The system then enhances the image. Compared with normal light reference image The similarity value between 0 and 1 is calculated from the input structural similarity function and then converted into the corresponding structural consistency loss term. The specific calculation process for the structural consistency loss is shown in the following formula: ; in, This represents the loss of structural consistency. This represents the structural similarity function. The specific calculations of the local window mean, variance, and covariance involved in the structural similarity function can be performed by those skilled in the art based on conventional image quality assessment standards. The underlying statistical logic is well-known in the field and will not be elaborated upon here.

[0063] Step S307, the system calculates the exposure consistency loss as described above. Reconstruction losses and structural consistency loss By performing linear combination, a total loss function is constructed to drive the entire neural network for backpropagation and parameter updates. The system assigns corresponding weight coefficients to each individual loss term to balance the contribution of physical constraints of different dimensions to the network optimization process. The specific calculation process of the overall loss function is shown in the following formula: ; in, Represents the overall loss function; Indicates the losses incurred during reconstruction; This represents the loss of structural consistency. This indicates a loss of exposure consistency; , and These represent the weighting coefficients for each loss item.

[0064] Step S308: To ensure that the gradient magnitudes provided by each loss term during backpropagation are within a reasonable and balanced range, the weight coefficients are adjusted. , and The value of is determined based on the specific convergence state of each loss. The weight coefficient is set within the range of real numbers greater than zero. In a specific embodiment, color reconstruction is the primary optimization objective. The value is set to 1.0; this is to preserve local high-frequency details in the image and constrain the macroscopic exposure level. The value is set within the range of [0.5, 1.0]. The value is set within the range of [0.1, 0.5]. The system calculates the overall loss function. Gradient descent optimization is performed to ensure that the network simultaneously meets multiple constraints such as pixel-level alignment, structural consistency, and physical exposure quantization mapping during iterative training.

[0065] After constructing the overall loss function, the system performs the construction of the joint optimization objective and the updating of model parameters. The specific operations can be divided into the following steps.

[0066] In step S309, the system determines the set of learnable parameters that need to be updated in the entire low-light image enhancement network. This set of learnable parameters includes the convolutional layer weights and biases in the illumination exposure prediction module, as well as the network parameters within components such as the multilayer perceptron, attention mechanism, and feedforward network in the dual-conditional adaptive enhancement module. The system uses the aforementioned constructed overall loss function as the optimization objective, aiming to find the optimal network parameter distribution that minimizes the overall loss function across the entire training dataset through iterative training. The specific calculation process of the joint optimization objective is shown in the following formula: ; in, This represents the set of optimal network parameters obtained after optimization. This represents the set of all learnable parameters in the current network. This represents the total number of image samples in the training dataset; Indicates the index of the training sample; Indicates that for the first The overall loss function calculated for each training sample in a single forward propagation.

[0067] Step S310: The system configures an optimizer to drive network parameter updates. In a specific embodiment, the system employs an adaptive moment estimation optimizer to jointly optimize the network parameters. The system calculates the gradient information of each layer's parameters during backpropagation based on the overall loss function. The optimizer updates the parameters of the illumination exposure prediction module and the dual-condition adaptive enhancement module layer by layer based on this gradient information and the internal state variables. For the setting of specific hyperparameters within the optimizer, such as the first-order momentum decay rate, the second-order momentum decay rate, and the numerical stability constant to prevent the denominator from reaching zero, those skilled in the art can configure them conventionally based on the actual hardware's memory capacity and convergence requirements. The basic gradient descent and parameter update logic is well-known in the art and will not be elaborated upon here.

[0068] Step S311: The system sets the initial learning rate and the corresponding dynamic decay strategy for the optimization process. The learning rate controls the step size of each parameter update. To ensure rapid convergence of the model in the early stages of training and stability of fine-tuning near the optimal solution in the later stages of training, the initial learning rate is set to 1×10⁻⁶. 4 Up to 5×10 4 The system introduces a cosine annealing strategy to dynamically adjust the learning rate, causing it to decrease non-linearly with increasing training cycles, and smoothly decay to a lower limit value at the end of the set maximum training cycle. In a specific embodiment, this lower limit value is set to 1×10⁻⁶. 6 .

[0069] In step S312, the system extracts low-light images and corresponding normal-light reference images from the training dataset in batches according to a preset batch size, and repeatedly performs the iterative optimization process of forward feature extraction, overall loss calculation, backpropagation differentiation, and network parameter update. When the system detects that the overall loss value of the model on the validation set tends to stabilize and no longer decreases within several consecutive epochs, or when the current training epoch reaches the preset maximum training epoch threshold, the system determines that the model has converged and ends the training process. The system then sets the optimal network parameters at this point. The model is then permanently stored and trained to enable it to perform adaptive image enhancement based on physical lighting conditions.

[0070] Specific application examples: To further verify the effectiveness and feasibility of the physical-guided low-light image enhancement method based on joint illumination and exposure modeling proposed in this invention, a specific application example is provided below.

[0071] In this embodiment, the low-light image enhancement system is deployed based on a deep learning framework. The hardware configuration includes a graphics processor (such as an NVIDIA RTX 3090 with 24GB of video memory), a multi-core processor (such as an Intel Core i7), and 32GB of memory.

[0072] During the model training phase, the system sets the following specific parameters: The training samples consist of low-light and normal-light image pairs. Data augmentation was performed on the samples using random rotation and flipping operations, and the samples were randomly cropped to a height... ,width The image patch. That is, the input low-light image. Compared with normal light reference image The tensor dimensions are all 3×128×128.

[0073] The system sets the batch size to 16 according to step S312. The optimizer, based on step S310, employs an adaptive moment estimation optimizer, with its first-order momentum decay rate set to 0.9 and its second-order momentum decay rate set to 0.999; the total number of training iterations is set to 6 × 10^6. 5 Second-rate.

[0074] The learning rate is configured according to step S311, with the initial learning rate set to 3×10. 4 A cosine annealing strategy is employed to gradually decay the value to a lower limit of 1×10⁻⁶ over the training period. 6 .

[0075] The overall loss function weights are set according to step S308: the weighting coefficients of the reconstruction loss. Weighting coefficients of structural consistency loss Weighting coefficients of exposure consistency loss .

[0076] The system is trained and tested on standard low-light benchmark datasets, including LOL-v1 (Low-Light dataset version 1, containing 485 pairs of training samples and 15 pairs of test samples) and LOL-v2-Real (real-scene low-light dataset version 2, containing 689 pairs of training samples and 100 pairs of test samples).

[0077] The evaluation metrics adopted three full-reference image quality assessment standards: PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity Index Measure), and LPIPS (Learned Perceptual Image Patch Similarity).

[0078] To verify the superiority of the method of this invention, it is compared with 19 existing mainstream methods. Table 1 shows the quantitative evaluation results of different low-light image enhancement methods.

[0079] It should be noted that in all the following tables: (1) "↑" indicates that the larger the value of the evaluation index, the better the image quality, fidelity and similarity after model enhancement; (2) "↓" indicates that the smaller (lower) the value of the evaluation index, the better the image quality and the smaller the perceptual distortion after model enhancement; (3) "CNN / GAN / Transformer / Mamba" represents the basic network architecture category used by the existing comparison methods (convolutional neural network, generative adversarial network, Transformer architecture and Mamba state space architecture, respectively). (4) “Ours” refers to the low-light image enhancement method proposed in this invention.

[0080] Meanwhile, the English names of the comparative methods mentioned in Table 1 and the specific existing method types they represent are explained as follows: MF: A traditional enhancement method based on multi-feature / multi-exposure fusion (Multi-Fusion); LIME: Low-light Image Enhancement via Illumination Map Estimation (LIME) is a traditional enhancement method based on illumination map estimation. BIMEF: A traditional enhancement method based on bio-inspired multi-exposure fusion. RetinexNet: A deep convolutional neural network based on Retinex theory; KinD: A practical low-light enhancement network for illumination and reflection decomposition and reconstruction (Kindling the Darkness). Z_DCE: Zero-reference Deep Curve Estimation method. RUAS: A method based on Retinex unrolling with architecture search. ELGAN: EnlightenGAN, an unsupervised generative adversarial network-based enhancement model. SCI: Self-Calibrated Illumination. DRBN-SKF++: Deep recursive bandgap network with selective kernel fusion; Restormer: A high-efficiency, high-resolution image restoration network based on Transformer; LLFormer: A Transformer network specifically designed for low-light vision (Low-Light VisionTransformer). RetinexMamba: A low-light enhancement method combining Retinex theory and the Mamba architecture; GDP: A method based on Generative Diffusion Prior; Diff-Retinex: A method combining diffusion models and Retinex theory; CLE: Contrastive Learning for Enhancement; GSAD: Generative Structure-Aware Diffusion. DiffDark / LatentDark: These refer to the diffusion model for dark-light images and the dark-light diffusion model based on the latent variable space, respectively.

[0081] Table 1. Quantitative evaluation results of different low-light image enhancement methods on the LOL-v1 and LOL-v2-Real datasets. As shown in Table 1, traditional methods generally perform poorly across various metrics due to their reliance on manually designed features. While some methods based on deep learning and diffusion models perform well on individual metrics, they suffer from insufficient structure preservation capabilities. In contrast, the method presented in this invention demonstrates the best overall performance, achieving a maximum PSNR of 24.833 and an SSIM of 0.863 on LOL-v1, as well as a minimum LPIPS of 0.093. This is primarily attributed to the invention's effective coordination of global enhancement intensity and local spatial guidance through illumination prediction and dual-conditional adaptive enhancement.

[0082] To further analyze the actual utility of each component within the system of this invention, this embodiment conducted ablation experiments with controlled variables on the LOL-v1 dataset. "Model-1" to "Model-7" represent model variants constructed after artificially reducing the complete network structure or loss function of this invention.

[0083] Table 2 shows the ablation experiment results of the system's internal network architecture module.

[0084] Table 2. Ablation experiment results of network architecture modules (LOL-v1 dataset) Referring to Table 2: For the dual-condition adaptive enhancement module, compared with Model-1 and the complete present invention, the lack of local spatial guidance leads to over-enhancement in bright areas and under-enhancement in dark areas, resulting in a PSNR decrease of 0.522dB; compared with Model-2, the lack of a global exposure benchmark causes macroscopic enhancement intensity to run out of control, resulting in a PSNR decrease of 0.383dB. For the illumination exposure prediction module, compared with Model-5, even when using local and global information simultaneously, using independent networks for parallel estimation leads to feature conflicts (PSNR of 24.441); the present invention adopts a feature pooling serial derivation mechanism, structurally ensuring the consistency of local and global physical measures.

[0085] Table 3 shows the ablation experiment results of the overall loss function of the system construction.

[0086] It should be noted that in Table 3: (1) "√" indicates that the model variant in the corresponding row used or included the loss function corresponding to that column for parameter optimization calculation during the network training phase; (2) "-" indicates that the model variant in the corresponding row did not include or removed the loss function corresponding to that column during the network training phase; (3) In the table header , , These correspond to the reconstruction loss, structural consistency loss, and exposure consistency loss constructed in the previous steps, respectively.

[0087] Table 3. Ablation experiment results of the overall loss function (LOL-v1 dataset) Referring to Table 3, when exposure consistency loss is removed from the system... At time (i.e., in the row where Model-6 is located), due to the lack of physical exposure ratio mapping constraints in the logarithmic domain, the brightness of the generated image drifts, and the PSNR drops to 24.481; further removal of structural consistency loss... At time (i.e., in the row where Model-7 is located), the network relies solely on reconstruction loss. This results in blurring of the image at edges and in textured areas, causing the SSIM to drop to 0.849. This invention, by combining the above three losses into a single overall loss function, successfully achieves optimal comprehensive enhancement across the three dimensions of pixel-level reconstruction, structural feature preservation, and physical exposure mapping accuracy.

Claims

1. A physically guided low-light image enhancement method based on joint modeling of illumination and exposure, characterized in that, Includes the following steps: The low-light image to be processed is input into the illumination and exposure prediction module, and the low-light image to be processed is subjected to feature extraction to generate a local illumination map. The global exposure factor is then derived serially from the local illumination map. The low-light image to be processed, the local illumination map, and the global exposure factor are input into the dual-condition adaptive enhancement module. The global exposure factor is combined with intensity modulation, and the local illumination map features are combined with spatial guided interactive operation. After feature reconstruction, the enhanced image is output. During the model training phase, a normal light reference image is acquired. Reconstruction loss and structural consistency loss are calculated for the output enhanced image and the normal light reference image. Exposure consistency loss is calculated for the derived global exposure factor. An overall loss function is constructed, and the optimizer is driven to jointly optimize the network parameters based on the overall loss function.

2. The physically guided low-light image enhancement method based on joint illumination and exposure modeling according to claim 1, characterized in that, The steps of extracting features from the low-light image to be processed to generate the local illumination map specifically include: Multi-scale features are captured from the input low-light image to be processed using a backbone architecture, and these multi-scale features are used as input features for the current processing stage. The extracted input features are transmitted to the illumination prediction block, and the illumination prediction block is used to perform layer normalization on the input features. The features that have completed layer normalization are used to generate query features, key features and value features through independent linear transformation operations. Self-attention calculation is performed using the query features, the key features and the value features. The query features, key features, and value features that have completed self-attention calculation are input into a feedforward network for nonlinear mapping processing to obtain the output features of the illumination prediction block; The illumination prediction block is stacked into four layers. The output features of the stacked illumination prediction block are mapped using convolutional layers to generate multi-channel features. The multi-channel features are compressed in the channel dimension to obtain single-channel features. The single-channel features are then output as the local illumination map.

3. The physically guided low-light image enhancement method based on joint modeling of illumination and exposure as described in claim 1, characterized in that, The steps of sequentially deriving the global exposure factor from the local illumination map features specifically include: The local illumination map features output from the end of the network are input into the exposure derivation block. The local illumination map features are downsampled in spatial dimension and compressed in channel dimension using multiple convolutional layers in the exposure derivation block. A linear rectified function is used to perform nonlinear mapping on the features output by the convolution between adjacent convolutional layers. After completing multi-layer convolution and non-linear mapping, pooling operation is used to compress the spatial dimension of the feature map and extract the feature vector representing global information. The feature vector is mapped to a single scalar value, and the scalar value is input into the activation function. The range of the input value is limited to a preset range. After the activation function mapping process, the global exposure factor in real number form is output.

4. The physically guided low-light image enhancement method based on joint modeling of illumination and exposure as described in claim 1, characterized in that, The specific steps of intensity modulation based on the global exposure factor include: The multilayer perceptron built inside the dual-condition adaptive enhancement module is used to perform nonlinear mapping on the global exposure factor to generate a first set of affine transformation parameters and a second set of affine transformation parameters. The low-light image to be processed is subjected to feature extraction to obtain low-light input features. The first set of affine transformation parameters is extracted and multiplied in the feature dimension to adjust the response amplitude of the features. At the same time, another parameter in the first set of affine transformation parameters is used to add to the low-light input features after multiplication to provide the offset of the feature reference value, thereby generating low-light modulation features.

5. The physically guided low-light image enhancement method based on joint illumination and exposure modeling according to claim 4, characterized in that, The steps for performing spatially guided interactive operations based on the local illumination map features specifically include: The local illumination map is used as a feature of the local illumination map in the calculation. Layer normalization operation is performed on the low light modulation feature. The feature is mapped to the feature subspace by an independent linear transformation layer and combined with dimension reshaping operation to generate query feature, key feature and value feature respectively. Calculate the product of the transpose matrix of the key feature and the matrix of the query feature, introduce a scaling factor to adjust the product result, and use a normalized exponential function to map the adjusted result into a normalized spatial guided weight. The local illumination map features are input into a convolutional layer for feature preprocessing. The spatial guidance weights are applied to the product of the convolutionally processed local illumination map features and the value features. The interactive features are added to the low-light modulation features through residual connections to output the spatial guidance features.

6. The physically guided low-light image enhancement method based on joint illumination and exposure modeling according to claim 5, characterized in that, The specific steps for outputting the enhanced image after feature reconstruction include: A layer normalization operation is performed on the spatial guidance features. The normalized spatial guidance features are then input into a feedforward network for nonlinear feature reshaping. The output of the feedforward network is then added to the original spatial guidance features to generate refined features. Extract the second set of affine transformation parameters generated by the multilayer perceptron mapping, the second set of affine transformation parameters including reconstruction scaling factor and reconstruction offset factor; The feature response intensity in the reconstruction stage is controlled by multiplying the feature dimensions of the refined features using the reconstruction scaling factor, and the feature reference value is calibrated by adding the refined features after the multiplication operation using the reconstruction offset factor to generate reconstruction intensity features through secondary calibration of the feature reference value.

7. The physically guided low-light image enhancement method based on joint illumination and exposure modeling according to claim 6, characterized in that, The step of outputting the enhanced image after feature reconstruction also includes: The generated reconstructed intensity features are transmitted to the projection layer at the output end, and a convolutional layer is used to project the multidimensional feature tensor back to the image color space in the channel dimension. After the projection layer outputs, an activation function is connected to limit the output value to a normalized range of image pixel values, and the enhanced image is finally output.

8. The physically guided low-light image enhancement method based on joint illumination and exposure modeling according to claim 1, characterized in that, The steps for calculating the exposure consistency loss based on the derived global exposure factor specifically include: The average exposure values ​​of the low-light image to be processed and the normal light reference image are calculated in the global dimension, respectively. The average exposure value is obtained by summing all pixel values ​​in the color channel dimension and the two-dimensional space dimension of the image and averaging them with the total number of pixels. The true exposure difference is calculated based on the extracted global exposure mean. The global exposure mean is then transformed into logarithmic space to obtain the difference. A smoothing constant is introduced into the input term of the logarithmic function. The derived global exposure factor is compared with the actual exposure difference obtained in logarithmic space, and the absolute difference is calculated using the norm to construct the exposure consistency loss.

9. The physically guided low-light image enhancement method based on joint modeling of illumination and exposure according to claim 1, characterized in that, The steps of calculating the reconstruction loss and the structural consistency loss for the output enhanced image and the normal light reference image, and constructing the overall loss function, specifically include: The reconstruction loss is calculated by using the norm to perform absolute difference between pixels at corresponding spatial locations in the enhanced image and the normal light reference image; The enhanced image and the normal light reference image are input into the structural similarity function to calculate the similarity value, and then converted into the structural consistency loss. The exposure consistency loss, reconstruction loss, and structural consistency loss are linearly combined, and corresponding weighting coefficients are set for each individual loss to construct the overall loss function.

10. The physically guided low-light image enhancement method based on joint illumination and exposure modeling according to claim 1, characterized in that, The steps for jointly optimizing network parameters based on the overall loss function-driven optimizer specifically include: An adaptive moment estimation optimizer is used to jointly optimize the network parameters using the overall loss function; An initial learning rate and a corresponding dynamic learning rate decay strategy are set. A cosine annealing strategy is introduced to dynamically adjust the initial learning rate so that the learning rate exhibits a non-linear decreasing trend as the training cycle increases and smoothly decays to the lower limit value at the end of the set maximum training cycle.

Images