A low-light enhanced image quality evaluation method and device, and a computer device

Abstract

The present application relates to computer vision technology, in particular to a kind of low light enhancement image quality evaluation method, device, computer equipment, method includes according to image illumination component multi-scale illumination feature map;From image extraction multistage original image feature;Illumination feature map is adjusted to the same size of corresponding original image feature by bilinear interpolation, injects each level original image feature by element multiplication, and calculates the difference of each level feature of original image and its enhanced image;Respectively from each level difference extraction texture information, fuse each level texture information;Respectively from each level difference extraction spatial information, fuse each level texture information;Fusion information is respectively input into classifier to obtain label, two labels are used as a pair of labels to input Bradley-Terry model and converted into unified global quality score.The present application makes model learn the perception difference of a pair of enhanced images and outputs label, simulates human preference selection process using Bradley-Terry model, so that the consistency of evaluation result and human subjective judgment is significantly improved.

Description

Technical Field

[0001] This invention relates to computer vision technology, and in particular to a method, apparatus, and computer device for evaluating the quality of low-light enhanced images. Background Technology

[0002] Existing technologies offer many ways to enhance low-light images, but the enhanced images still cannot match human subjective perception.

[0003] Several representative methods exist for objective quality assessment of low-light enhanced images. Yang et al. proposed the BEHN metric, which integrates enhancement perception, structure preservation, and color naturalness features using AdaBoost-RF; Zhai et al. proposed the LIEQA model, which improves adaptability to low-light environments with the help of a four-dimensional evaluation system; Wang et al. designed a pairwise learning network that extracts patterns and fuses semantics using Intrinsic Perception Feature Extraction (INPFE) and Impairment Perception Feature Extraction (IMPFE) modules, achieving better performance than traditional metrics in complex night scenes through human-aligned preference ranking. However, these methods have significant shortcomings: most rely on enhanced images for feature extraction, failing to fully utilize key information such as illumination; they employ simple feature fusion strategies, limiting feature representation capabilities and thus limiting model performance; and most methods assign absolute scores to each enhanced image, failing to fully simulate the human "either / or" preference mechanism. Summary of the Invention

[0004] To address the problems of underutilization of illumination information, simplistic feature fusion strategies, and difficulty in simulating human preference mechanisms in existing objective quality assessment methods for low-light enhanced images, this invention proposes a new low-light enhanced image quality assessment method. Based on multi-level illumination injection and hierarchical difference perception, this method assesses whether the enhanced image of the original low-light image is more suitable for human visual perception than the original low-light image. Specifically, it includes the following steps:

[0005] Based on Retinex theory, a dual-branch network is used to decompose the input image into illumination and reflection components respectively;

[0006] The lighting components are input into baseNet for basic feature extraction, and then multi-scale lighting feature maps are generated through pyramid pooling.

[0007] The image input is processed through a pre-trained MambaOut network to extract features at multiple levels, resulting in multi-level original image features.

[0008] The illumination feature map is adjusted to the same size as the feature map of the corresponding layer of the MambaOut network through bilinear interpolation, and then the original image features at each level are injected through element-wise multiplication.

[0009] Obtain the original low-light image and its enhanced image at each level of original image features, and calculate the differences between the features at each level;

[0010] Texture information is extracted from each high-resolution differential feature through a spatial-frequency domain sensing branch, and then the extracted texture information is fused through a cross-scale fusion module; semantic information is extracted from each low-resolution differential feature through a semantic sensing module, and then the extracted semantic information is fused through a cross-scale fusion module.

[0011] The features obtained from cross-scale fusion are input into the classification system to obtain paired labels. The paired labels are then input into the Bradley-Terry model to convert them into a unified global quality score.

[0012] Compared with the prior art, the present invention has the following beneficial effects:

[0013] 1. This invention employs multi-level illumination injection, which hierarchically injects illumination information into the feature extraction process, thereby enhancing the role of brightness information in evaluation. Hierarchical difference perception then targets high-level semantic features and low-level texture features respectively. Furthermore, cross-scale fusion enables refined utilization of semantic and texture features, solving the problems of simple feature fusion and insufficient utilization of key information in existing methods. This improves the model's adaptability to complex low-light scenes, as low-light image acquisition is prone to degradation issues, and different low-light conditions present diverse noise patterns, different illumination levels, and complex scene content. This full utilization of features allows the model to better cope with these complex situations.

[0014] 2. This invention uses a pairwise comparison design to allow the model to learn the perceptual differences between pairs of enhanced images and output preference labels. Then, the Bradley-Terry (BT) model is used to convert these pairwise labels into a unified global quality score, simulating the human "either / or" preference selection process, rather than assigning an absolute score to each enhanced low-light image as most existing methods do. This approach is more in line with human subjective judgment patterns, significantly improving the consistency between the evaluation results and human subjective judgment. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a low-light enhancement image quality assessment method according to the present invention.

[0016] Figure 2 This is a schematic diagram of the multi-scale measurement feature injection process of the present invention;

[0017] Figure 3 This is a schematic diagram of the data processing of the hierarchical sensing module of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] This invention proposes a specific implementation method for low-light enhanced image quality assessment, which includes the following steps:

[0020] S1. Based on Retinex theory, a dual-branch network is used to decompose the input image into illumination and reflection components. The reflection branch retains high-frequency texture details through skip connections between U-Net-style encoder / decoder blocks and accurately reconstructs surface reflectivity through multi-level feature fusion using bilinear interpolation and element-wise addition. The illumination branch, through a cascaded upsampling path, goes through consecutive 3×3 convolutions and bilinear upsampling stages. Since there are no skip connections, it forms a smooth illumination map that conforms to physical priors.

[0021] S2. In the feature extraction process, the input illumination map is first subjected to basic feature extraction using BaseNet, followed by pyramid pooling to form a feature map. The original image, on the other hand, undergoes multi-level feature extraction using a pre-trained MambaOut network. This network sequentially passes through initial convolutional downsampling to obtain high-resolution texture features, and then intermediate units that progressively downsample and expand the channel dimensions. Subsequently, the illumination components are hierarchically injected into the feature map of the MambaOut network through multi-level fusion, constructing a feature pyramid from local to global features. This enhances the model's sensitivity to brightness changes and addresses the problem that most existing methods rely on enhancing low-light images for feature extraction, failing to fully utilize available key information such as illumination.

[0022] S3. For paired image enhancement, the extracted features are processed hierarchically. For features related to high-level semantics, layer-by-layer differences are calculated using a channel alignment function. The resulting difference features are fed into the semantic branch. Here, channel attention weights and a global context encoder capture semantic-level differences. This module reweights key channel features and uses a continuous Transformer to model long-range dependencies, focusing on semantic content changes in image pairs, such as systematic differences in object attributes or scene structure changes. For features related to low-level texture information, a dedicated spatial-frequency branch is developed to jointly process spatial and frequency domain information. In frequency domain analysis, a two-dimensional Fourier transform is applied to the difference features to extract the real part spectrum, and depthwise separable convolution is used to dynamically modulate the frequency response to suppress high-frequency noise. In spatial domain analysis, spatial attention is used to enhance local structure-sensitive regions. This is achieved by convolving the average pooling and max pooling results of the difference features through channel concatenation, followed by a Sigmoid activation function. By performing nonlinear fusion of the extracted dual-path features from the input image, fine-grained differential responses are generated. Then, these features are fused through cross-scale adaptive weighting to produce preference decisions and output pairwise preference labels. This solves the problem that the simple fusion strategy in existing methods limits feature expression and leads to limited model performance.

[0023] This embodiment proposes a specific implementation method for low-light enhanced image quality assessment, including a complete process of illumination decomposition, multi-level illumination injection, hierarchical difference perception, and global score generation. The network structure parameters and operations involved are all reproducible specific settings. Figure 1 In this embodiment, the original low-light image I j Enhanced image I of the original low-light image j The input illumination decomposition network yields the corresponding illumination components, i.e., the original low-light image I. j Corresponding illumination component L j Enhanced image I of the original low-light image j Corresponding illumination component L i , will (I j ,L j ) and (I i ,L i Two weight-shared multi-scale brightness feature injection modules are input respectively to calculate the difference between the two images at each scale. The high-resolution features are input into the spatial-frequency domain perception branch and the low-resolution features are input into the semantic perception branch. The outputs of the two branches are input into the classifier to obtain two labels. The obtained labels are input into the model to obtain the quality score of the enhanced image. The following embodiment will describe this process with steps 1 to 4.

[0024] Step 1: Illumination decomposition.

[0025] Based on Retinex theory, a dual-branch network is used to decompose the illumination and reflection components of the input low-light enhancement image. The specific structure is as follows:

[0026] The reflection branch adopts a U-Net-style symmetric encoder-decoder structure. The encoder contains four convolutional blocks (each block consists of two 3×3 convolutional layers with a stride of 1, padding of 1, and ReLU activation function). The decoder is symmetrical to the encoder, and the i-th layer of the encoder and the (4-i)-th layer of the decoder are connected via skip connections (when the feature map sizes are inconsistent, bilinear interpolation is used to adjust them to the same size before element-wise addition) to achieve multi-level feature fusion. The reflection component is output through this branch. H and W are the height and width of the input image. In this embodiment, the default size is 224×224.

[0027] The illumination branch employs a cascaded upsampling path, comprising three upsampling units. Each upsampling unit consists of a 3×3 convolutional layer (stretch 1, padding 1, ReLU activation) and a bilinear upsampling layer (with a scaling factor of 2). There are no skip connections to ensure the smoothness of the illumination map. This branch outputs the illumination components. ;

[0028] The above decomposition process satisfies the condition that the input image equals the illumination component. Reflective components, This is element-wise multiplication. Step 2: Multi-level illumination injection.

[0029] This embodiment uses image I i Taking the example of this to illustrate the multi-level lighting injection process, the lighting components obtained in step 1 are... The hierarchical injection feature extraction process is as follows:

[0030] From lighting components Extracting illumination features, including: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] The baseNet is input for basic feature extraction. The baseNet consists of two 3×3 convolutional layers (stretch 1, padding 1) and one max pooling layer (2×2, stretch 2). Then, multi-scale illumination feature maps are generated through pyramid pooling. Pyramid pooling includes four pooling scales: 1×1, 2×2, 4×4, and 8×8. After pooling at each scale, the channels are compressed to 64 through a 1×1 convolution to obtain the injected feature map.

[0031] This invention extracts original image features from the input image. It employs a pre-trained MambaOut network for feature extraction. In this embodiment, the MambaOut network uses a four-level feature extraction process, specifically including:

[0032] MambaOut stage 1 consists of a single 3×3 convolutional layer (stretch 2, padding 1). MambaOut stage 1 has 64 output channels. Image I is extracted using MambaOut stage 1. i The features are determined by using injection head 1 to match the size of the injected feature map with the feature map obtained through MambaOut stage 1, and then performing a Hadamard product on the two feature maps to obtain image I. i Level 1 features;

[0033] MambaOut stage 2 includes a 3×3 convolutional layer (stretch 2, padding 1). The channels of MambaOut stage 2 are expanded to 128, utilizing image I. i The first-level features are used as input to MambaOut stage 2. Injection head 2 is used to match the size of the injected feature map with the feature map obtained through MambaOut stage 2. Then, the two feature maps are subjected to Hadamard product to obtain image I. i Level 2 features;

[0034] MambaOut stage 3 includes a 3×3 convolutional layer (stretch 2, padding 1). The channels of MambaOut stage 2 are expanded to 256, utilizing image I. i The second-level features are used as input to MambaOut stage 3. Injection head 3 is used to match the size of the injected feature map with the feature map obtained through MambaOut stage 3. Then, the two feature maps are subjected to Hadamard product to obtain image I. i Level 3 features;

[0035] MambaOut stage 4 includes a 3×3 convolutional layer (stretch 2, padding 1). The channels of MambaOut stage 2 are expanded to 512, utilizing image I. i The third-level features are used as input to MambaOut stage 4. Injection head 4 is used to match the size of the injected feature map with the feature map obtained through MambaOut stage 4. Then, the two feature maps are subjected to Hadamard product to obtain image I. i Level 4 features;

[0036] As an optional implementation, in this embodiment, the injection head adjusts the illumination feature map to the same size as the corresponding level feature map of MambaOut using bilinear interpolation, and then injects it into the 1st to 4th level feature maps through element-wise multiplication to obtain image I. i Multi-level image features , Image I i The k-th level feature; similarly, image I can be obtained. j Multi-level image features , Image I jThe k-th level feature.

[0037] Step 3: Hierarchical difference perception.

[0038] The input of this invention is the original low-light image I. j and its enhanced diagram I i The image pairs formed by the images are processed through steps 1 and 2 to obtain image I. j Multi-level image features and image I i Multi-level image features Perform stratified difference processing, such as Figure 3 ,include:

[0039] First, the images are divided into a high-resolution set and a low-resolution set, each set including at least one image feature. Those skilled in the art perform the set division according to the resolution. In this embodiment, four-level feature extraction is used, with the first and second level image features assigned to the high-resolution set and the third and fourth level image features assigned to the low-resolution set.

[0040] For the features of the high-resolution set, the difference of each level of features is calculated separately, i.e. (k={1,2}) Represents the original low-light image Its corresponding enhanced image The difference in the k-th level features between them;

[0041] right (k={1,2}) performs a two-dimensional Fourier transform to extract the real part spectrum, and obtains the frequency domain features by modulating the frequency response through a 3×3 depth separable convolution with 128 output channels;

[0042] right (k={1,2}) undergo 2×2 average pooling and max pooling. The pooling results are concatenated along the channel dimension and then used to generate spatial attention weights through a 3×3 convolution with 1 output channel and sigmoid activation. Spatial attention weights and Perform element-wise multiplication to obtain the spatial domain characteristics;

[0043] After unifying the frequency domain and spatial domain features through 1×1 convolution and then adding the elements together, texture difference features are obtained.

[0044] The fusion is achieved through adaptive weights (calculated based on feature variance, with a weight sum of 1). , The obtained features are then fed into a binary classifier (a 2-layer multilayer perceptron MLP with a 2-dimensional output) to generate the first preference label. (when A value of 1 indicates the original low-light image. Superior quality, when A value of 2 indicates that the enhanced image corresponds to the original low-light image. Superior quality;

[0045] For the characteristics of low-resolution sets, alignment is performed using a channel alignment function, and then the differences are calculated. The alignment function used in this embodiment is... A 1×1 convolution is used to unify the channels of the feature map to 256 dimensions before calculating the differences. (k={3,4});

[0046] right Global average pooling is performed on k={3,4}. A 256-dimensional weight vector is generated using a cascaded multilayer perceptron (MLP) with 64-dimensional hidden layers and sigmoid activation. This weight vector is then used in conjunction with... Perform element-wise multiplication (k={3,4}) to obtain the weighted features;

[0047] We use a cascaded two-layer Transformer (each layer includes 8 attention heads and 512 hidden dimensions) to extract long-term dependencies from weighted features to obtain semantic differential features;

[0048] The fusion is achieved through adaptive weights (calculated based on feature variance, with a weight sum of 1). , The obtained features are then fed into a binary classifier (a 2-layer multilayer perceptron MLP with a 2-dimensional output) to generate a second preference label. (when A value of 1 indicates the original low-light image I. j Superior quality, when A value of 2 indicates that the enhanced image I corresponds to the original low-light image. i Superior quality.

[0049] Step 4: Bradley-Terry model.

[0050] The Bradley-Terry model is a commonly used probabilistic model for comparing the relative capabilities or preferences of two or more items. In this model, each item is assigned a capability parameter representing its relative capability. By observing the comparison results between items, these capability parameters can be estimated and predictions made. For example, the BradleyTerry2 package developed by Tencent Cloud can handle binary comparison data, meaning each comparison result involves only two items. This invention inputs the first and second preference labels as a label pair into the Bradley-Terry (BT) model, transforming these paired labels into a unified global quality score. Specifically, for each scene, a preference matrix is ​​first constructed, where each entry represents the empirical probability that image i is superior to image j based on the collected comparison results. The BT model estimates the potential quality score by assuming that the preference probabilities follow a logical form. These scores are optimized using maximum likelihood estimation to ensure global consistency across all comparisons. Finally, the obtained scores are normalized to obtain the final quality assessment metric, thus simulating the human "either / or" preference mechanism. This replaces the traditional evaluation method of assigning absolute scores to each enhanced low-light image, solving the problem that existing methods fail to adequately simulate human preference mechanisms.

[0051] Tables 1 and 2 show the performance of this invention and other existing image quality assessment methods on the LE dataset and the RNTIEQA dataset. This embodiment evaluates the effectiveness of the model from two dimensions: prediction accuracy (Pearson Linear Correlation Coefficient, PLCC) and monotonicity (Spearman Rank-Order Correlation Coefficient, SRCC).

[0052] Table 1

[0053] SRCC PLCC NIQE 0.0632 0.0828 BRISQUE 0.4689 0.3169 CNNIQA 0.5211 0.5349 DBCNN 0.6726 0.6903 HyperIQA 0.7772 0.7851 VCRNet 0.6514 0.6526 StairIQA 0.7575 0.7718 VIPNet 0.6846 0.6708 TemPQT 0.7018 0.7115 PINet 0.7822 0.7822 AGAIQA 0.8127 0.8021 SGDNet 0.7817 0.7825 This invention 0.8144 0.8208

[0054] Table 2

[0055] SRCC PLCC NIQE 0.0094 0.0061 IL-NIQE 0.0245 0.0162 NI-QMC 0.0576 0.0377 SNP-NIQE 0.1333 0.088 PIQE 0.2058 0.1376 NUIQ 0.3826 0.2656 NBIQA 0.4342 0.3008 BlIInds-II 0.4391 0.3041 BRISQUE 0.464 0.3209 FRIQUEE 0.475 0.3296 CNNIQA 0.5425 0.3851 VCRNet 0.6308 0.4513 DBCNN 0.6491 0.4684 This invention 0.6696 0.6722

[0056] Among them, the Natural Image Quality Evaluator (NIQE) comes from the paper "Making a “Completely Blind” Image Quality Analyzer"; the BRISQUE blind / no-reference image spatial quality evaluator comes from the paper "BLIND / REFERENCELESS IMAGE SPATIAL QUALITY EVALUATOR"; the CNNIQA convolutional neural network for no-reference image quality assessment comes from the paper "Convolutional Neural Networks for No-Reference Image Quality Assessment"; the application of the deep bilinear convolutional neural network DBCNN in image quality assessment comes from the paper "Blind Image Quality Assessment Using a Deep Bilinear Convolutional Neural Network"; the HyperIQA quality assessment method comes from the paper "Blindly Assess Image Quality in the Wild Guided by A Self-Adaptive Hyper Network"; and the VCRNet deep learning model for no-reference image quality assessment comes from the paper "VCRNet: Visual CompensationRestoration Network for No-Reference Image Quality". StairIQA is an image quality assessment tool; VIPNet, an image quality assessment method, comes from the paper "Visual Interaction Perceptual Network for Blind Image Quality Assessment"; TemPQT, an image quality assessment method, comes from the paper "Blind Image Quality Assessment via Transformer Predicted Error Map and Perceptual Quality Token"; AGAIQA comes from the paper "Blind Image Quality Assessment via Adaptive Graph Attention".The image quality assessment methods SGDNet, IL-NIQE, and NBIQA are derived from the paper "Blind image quality index with high-level Semantic Guidance and low-level fine-grained representation". NI-QMC comes from the paper "No-Reference Quality Metric of Contrast-Distorted Images Based on Information Maximization". SNP-NIQE is from the paper "Unsupervised Blind Image Quality Evaluation via Statistical Measurements of Structure, Naturalness, and Perception". PIQE is from the paper "BLIND IMAGE QUALITY EVALUATION USINGPERCEPTION BASED FEATURES". NUIQ is from the paper "Underwater ImageEnhancement Quality Evaluation: Benchmark Dataset and Objective Metric". The image quality assessment method Blinds-II is based on the paper "Blind Image Quality Assessment: A Natural Scene Statistics Approach in the DCTDomain"; the image quality assessment method BRISQUE is based on the paper "BLIND / REFERENCELESS IMAGE SPATIALQUALITY EVALUATOR".The image quality assessment method FRIQUEE comes from the paper "Perceptual Quality Prediction on Authentically Distorted Images Using a Bag of Features Approach".

[0057] Tables 1 and 2 show the PLCC and SRCC metrics of our invention and other existing image quality assessment models on the LE and RNTIEQA datasets, respectively. As shown in Table 2, on the RNTIEQA dataset, our method achieves an SRCC of 0.6696 and a PLCC of 0.6722, while other methods such as NIQE achieve only 0.0094 and 0.0061 SRCC, BRISQUE achieves 0.464 and 0.3209 SRCC, and even the well-performing DBCNN only achieves 0.6491 and 0.4684 SRCC. As shown in Table 1, on the RNTIEQA dataset, our solution achieves an SRCC of 0.6696 and a PLCC of 0.6722, outperforming methods such as CNNIQA (SRCC 0.5425, PLCC 0.3851) and VCRNet (SRCC 0.6308, PLCC 0.4513), thus more accurately reflecting subjective quality.

[0058] Ablation experiments were conducted under consistent experimental settings, with the effectiveness of each module verified by sequentially removing or modifying specific components. When step 2 (i.e., multi-level illumination injection) was removed, raw feature extraction was performed directly using MambaOut, resulting in a slight decrease in SRCC and KRCC scores, while PLCC scores decreased significantly. When step 3 (hierarchical difference perception) was removed, the multi-stage features obtained from MambaOut were concatenated, then processed through global average pooling, and input into the MLP layer for preference label prediction. This resulted in a reduction in all metrics, demonstrating that these two steps played a crucial role in improving evaluation performance.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for evaluating the quality of low-light enhanced images, characterized in that, The evaluation of whether an enhanced image of the original low-light image is more suitable for human visual perception than the original low-light image, based on multi-level illumination injection and hierarchical difference perception, specifically includes the following steps: According to Retinex theory, the process of using a dual-branch network to decompose the input image into illumination and reflection components includes: Surface reflectivity is reconstructed using a U-Net module, in which skip connections are used between the encoder and decoder, and multi-level feature fusion is performed through bilinear interpolation and element-wise addition. A smooth illumination map is obtained by cascading upsampling paths, which consist of cascaded 3×3 convolutions and bilinear upsampling. The lighting components are input into baseNet for basic feature extraction, and then multi-scale lighting feature maps are generated through pyramid pooling. The image input is processed through a pre-trained MambaOut network to extract features at multiple levels, resulting in multi-level original image features. The illumination feature map is adjusted to the same size as the feature map of the corresponding layer of the MambaOut network through bilinear interpolation, and then the original image features at each level are injected through element-wise multiplication. Obtain the original low-light image and its enhanced image at each level of original image features, and calculate the differences between the features at each level; Texture information is extracted from each high-resolution difference feature through a spatial-frequency domain aware branch, including: Two-dimensional Fourier transforms are performed on each differential feature to extract the real part spectrum. The frequency domain features are obtained by outputting a 3×3 depth-separable convolutional modulation frequency response with 128 output channels. Each differential feature is subjected to 2×2 average pooling and max pooling respectively. After the pooling results are concatenated, they are sequentially passed through 3×3 convolution and Sigmoid activation of output channel 1 to generate spatial attention weights. The spatial attention weights are then used to weight the corresponding differential features to obtain spatial features. After unifying the frequency domain and spatial domain features through 1×1 convolution and then adding the elements together, texture difference features are obtained. The texture difference features corresponding to each difference feature are fused through an adaptive weight with a weight sum of 1 to obtain texture information; The extracted texture information is then fused using a cross-scale fusion module; Semantic information is extracted from each low-resolution differential feature using a semantic awareness module, including: Global average pooling is performed on the differential features at each level. After pooling, a weight vector is generated by a two-layer cascaded multilayer perceptron. The corresponding differential features are then weighted by element-wise multiplication using the weight vector. The weighted differential features are input into two cascaded Transformer layers to extract long-term dependency features as the semantic features of the current differential features; The semantic features of all differences are fused together using adaptive weights with a weight sum of 1 to obtain semantic information; The extracted semantic information is then fused using a cross-scale fusion module; The features obtained from cross-scale fusion are input into the classification system to obtain paired labels. The paired labels are then input into the Bradley-Terry model to convert them into a unified global quality score.

2. The low-light enhancement image quality assessment method according to claim 1, characterized in that, The baseNet consists of two 3×3 convolutional layers with a stride of 1 and padding of 1, and a 2×2 max pooling layer with a stride of 2, cascaded together. The pyramid pooling includes four pooling sizes cascaded in sequence: 1×1, 2×2, 4×4, and 8×8. After each pooling, the channels are compressed to 64 by a 1×1 convolution.

3. The low-light enhancement image quality assessment method according to claim 2, characterized in that, The MambaOut network consists of four cascaded 3×3 convolutional layers with a stride of 2 and padding of 1. The number of channels in the first to fourth levels are 64, 28, 256, and 512, respectively.

4. The low-light enhancement image quality assessment method according to claim 1, characterized in that, The calculation of the differences at each level of features includes: unifying the same level features of the original low-light image and the corresponding enhanced image of the original low-light image to 256 channels using a channel alignment function, and then using the difference between the same level features of the unified original low-light image and its enhanced image as the difference of that level of feature.

5. A low-light enhancement image quality assessment device, characterized in that, To implement the low-light enhancement image quality assessment method of claim 1, the method includes: A dual-branch network is used to decompose the input image into illumination and reflection components; The system employs a weight-sharing approach, comprising a first multi-scale metric feature injection and a second multi-scale metric feature injection. The first multi-scale metric feature injection processes the original low-light image, while the second multi-scale metric feature injection processes the enhanced image corresponding to the original low-light image. Each multi-scale metric feature injection includes a multi-scale illumination feature map extraction module and a pre-trained MambaOut network. The multi-scale illumination feature map extraction module includes baseNet and pyramid pooling. baseNet extracts basic features from the illumination components, and pyramid pooling generates multi-scale illumination feature maps based on the basic features. The pre-trained MambaOut network performs multi-level feature extraction on the input image to obtain multi-level original image features. The illumination feature map is then adjusted to the same size as the corresponding level feature map of the MambaOut network using bilinear interpolation. Element-wise multiplication is used to inject the original image features at each level, and the fused features at each level are used as the output of the multi-scale metric feature injection. The feature difference module is used to calculate the differences in various scale features between the original low-light image and its enhanced image; Spatial-frequency domain aware branch is used to extract texture information from high-resolution differences; A semantic awareness module is used to extract semantic information from low-resolution differences; A classifier is used to classify images based on their texture and semantic information, and to use the classification results of a set of images as paired labels. The Bradley-Terry model is used to convert pairwise labels into a uniform global quality score.

6. A computer device, characterized in that, The computer device includes: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the low-light enhancement image quality assessment method as described in claim 1.

Images