No-reference quality evaluation method and system for low-light enhanced image and model training method

Abstract

The invention relates to a non-reference quality evaluation method and system for a low-light enhancement image and a model training method. The method includes: converting an image to an HVI color space to separate luma and chroma components; performing multi-scale decomposition on the two components; extracting global features from the decomposed low-frequency components by adopting a self-attention-based module, extracting local features from the high-frequency components by adopting a convolutional neural network-based module, and realizing interaction enhancement of brightness and chrominance features under the same scale through a cross attention mechanism; and finally, progressively fusing the multi-scale features and carrying out regression to obtain a quality score. The invention also provides a system for realizing the method and a corresponding model training method. According to the method, coupling distortion of brightness and chromaticity in the low-light-enhanced image can be effectively decoupled and modeled, and evaluation of global illumination and local details is considered, so that the accuracy and robustness of non-reference quality evaluation are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of image processing, and in particular to a no-reference quality assessment method, system, and model training method for low-light enhanced images. Background Technology

[0002] With the deep integration of the Internet of Things (IoT) and artificial intelligence (AI), low-light images serve as a crucial data source for IoT front-end acquisition devices. Their quality significantly impacts the decision-making reliability of edge intelligent systems in critical tasks such as marine exploration, nighttime urban security, coal mine safety monitoring, and robot navigation and visual positioning. Compared to image acquisition under normal lighting conditions, images in low-light environments often suffer from uneven brightness, reduced contrast, and loss of detail due to insufficient illumination. Furthermore, sensors operating in high-gain mode introduce substantial random noise and color shift, further degrading image quality and affecting the robustness of backend perception algorithms and the user's visual experience.

[0003] To alleviate the aforementioned problems, low-light image enhancement (LLE) methods have been extensively studied and have become a major research hotspot in the field of computer vision. However, existing methods generally suffer from problems such as brightness overcompensation, color shift, and noise amplification, resulting in varying degrees of visual distortion in the enhancement results. Against this backdrop, constructing a reliable low-light image enhancement quality assessment model (LIEQA) is of great significance for optimizing enhancement algorithms, improving the visual quality experience for end users, and building quality feedback mechanisms for IoT systems.

[0004] Generally, image quality assessment (IQA) methods can be divided into three categories: full-reference (FR), reduced-reference (RR), and no-reference / blind (NR). In low-light image enhancement quality assessment applications, the practical applicability of FR and RR methods is limited due to the difficulty in obtaining high-quality reference images. NR methods, which rely solely on the image being evaluated, become a more feasible choice.

[0005] In recent years, although numerous general-purpose Blind Image Quality Assessment (BIQA) methods have emerged and achieved excellent performance in daytime natural image quality assessment tasks, these methods are primarily designed for natural scene images, and their features and training strategies may not be applicable to low-light enhanced images. The fundamental reason is that degradation and enhancement work together in low-light scenes, leading to higher complexity and structural anisotropy in image quality deterioration. This is not only affected by global illumination but also highly dependent on factors such as non-uniform local brightness distribution, color distortion, and the characteristics of enhancement algorithms. Currently, some BIQA methods for specific degradation scenarios share some similarities with the degradation characteristics of low-light enhancement. For example, contrast enhancement IQA mainly focuses on the loss of detail and visual unnaturalness caused by unreasonable brightness distribution, which corresponds to phenomena such as overexposure and contrast imbalance caused by brightness compensation in low-light enhanced images. However, such similar methods cannot fully identify and model more typical quality defects in low-light enhanced scenes, such as "false brightness" caused by over-enhancement, local highlight clipping, visual abruptness caused by color distortion, and texture blurring and perceptual unnaturalness caused by residual underexposed areas. Summary of the Invention

[0006] To achieve an accurate and robust no-reference objective assessment of the perceived quality of low-light enhanced images, this application provides a no-reference quality assessment method, system, and model training method for low-light enhanced images.

[0007] Firstly, the reference-free quality assessment method for low-light enhanced images provided in this application adopts the following technical solution:

[0008] A referenceless quality assessment method for low-light enhanced images includes the following steps:

[0009] Acquire the low-light enhancement image to be evaluated;

[0010] The low-light enhanced image is converted from the original color space to the HVI color space to separate the luminance and chrominance components;

[0011] The luminance component and the chrominance component are decomposed into multiple scales to obtain sets of luminance sub-components and sets of chrominance sub-components corresponding to multiple different scales.

[0012] At the same scale, hierarchical feature extraction and interactive processing are performed on the corresponding luminance sub-component and chrominance sub-component respectively to obtain enhanced luminance features and enhanced chrominance features at that scale after interactive enhancement; wherein, the hierarchical feature extraction and interactive processing includes using a self-attention-based module to process at least one low-frequency scale component, and using a convolutional neural network-based module to process other scale components.

[0013] The enhanced luminance features and enhanced chrominance features from all scales are fused to generate the final fused feature;

[0014] Based on the final fusion features, a quality assessment score for the low-light enhanced image is obtained through regression.

[0015] Preferably, the multi-scale decomposition is a Laplace pyramid decomposition.

[0016] Preferably, the number of decomposition scales of the Laplace pyramid is 3.

[0017] Preferably, the hierarchical feature extraction and interaction processing specifically includes:

[0018] Shallow feature extraction is performed on the luminance subcomponent and chrominance subcomponent at different scales to obtain shallow luminance features and shallow chrominance features. The shallow feature extraction includes using convolution operations with different receptive fields to aggregate multi-scale contextual information.

[0019] Deep feature extraction is performed on the shallow luminance features and shallow chrominance features respectively to obtain deep luminance features and deep chrominance features;

[0020] Deep luminance features and deep chrominance features at the same scale interact through a cross-attention mechanism to obtain enhanced luminance features and enhanced chrominance features.

[0021] Preferably, in the deep feature extraction step:

[0022] For at least one low-frequency scale component that represents global low-frequency information obtained through multi-scale decomposition, feature extraction is performed using a unit based on the Transformer architecture.

[0023] For other scale components that represent local high-frequency information obtained through multi-scale decomposition, feature extraction is performed using an enhanced residual unit based on a convolutional neural network and incorporating an attention mechanism.

[0024] Preferably, the interaction via the cross-attention mechanism specifically includes:

[0025] Using deep luminance features as query vectors and deep chrominance features as key and value vectors, calculate the first cross attention and use the result to enhance the deep luminance features to obtain enhanced luminance features;

[0026] Using deep chroma features as the query vector and deep luminance features as the key and value vectors, a second cross-attention is calculated, and the result is used to enhance the deep chroma features to obtain enhanced chroma features.

[0027] Preferably, the enhanced luminance features and enhanced chrominance features from all scales are fused to generate the final fused feature, including:

[0028] At each scale, based on the enhanced brightness features and enhanced chromaticity features of that scale, a fusion weight is adaptively generated, and the enhanced brightness features and enhanced chromaticity features of that scale are weighted and fused according to the fusion weight to obtain the fused features of that scale.

[0029] The fused features at different scales are aligned for resolution, and the high-resolution scale features are fused into the low-resolution scale features step by step in order from high-resolution scale to low-resolution scale. The fusion results are then refined to obtain the final fused features.

[0030] Preferably, the step of obtaining the quality assessment score based on the final fusion features through regression specifically includes:

[0031] The final fused features are subjected to global average pooling to obtain a feature vector;

[0032] The feature vector is input into a multilayer perceptron for nonlinear mapping, and the quality assessment score is output.

[0033] Secondly, the reference-free quality assessment system for low-light enhanced images provided in this application adopts the following technical solution:

[0034] A referenceless quality assessment system for low-light enhanced images includes:

[0035] The image acquisition module is configured to acquire the low-light enhanced image to be evaluated;

[0036] The color space conversion module is configured to convert the low-light enhanced image from the original color space to the HVI color space to separate the luminance component and the chrominance component.

[0037] The multi-scale decomposition module is configured to perform multi-scale decomposition on the luminance component and the chrominance component respectively, to obtain a set of luminance sub-components and a set of chrominance sub-components corresponding to multiple different scales.

[0038] The hierarchical feature extraction and interaction module is configured to perform hierarchical feature extraction and interaction processing on the corresponding luminance sub-component and chrominance sub-component at the same scale, respectively, to obtain enhanced luminance features and enhanced chrominance features at that scale after interaction enhancement; wherein, the hierarchical feature extraction and interaction processing includes using a self-attention-based sub-module to process at least one low-frequency scale component, and using a convolutional neural network-based sub-module to process other scale components.

[0039] The feature fusion module is configured to fuse enhanced luminance features and enhanced chrominance features from all scales to generate the final fused feature;

[0040] The quality regression module is configured to regress a quality assessment score of the low-light enhanced image based on the final fusion features.

[0041] Thirdly, the method for training a no-reference quality assessment model for low-light enhanced images provided in this application adopts the following technical solution:

[0042] A method for training a no-reference quality assessment model for low-light enhanced images includes:

[0043] Obtain a training dataset containing multiple low-light enhanced images and their corresponding subjective quality scores;

[0044] Construct a quality assessment model, which is configured to perform the aforementioned no-reference quality assessment method;

[0045] Using the low-light enhanced image as input, the subjective quality score as the supervision label, and the mean absolute error as the loss function, the model is trained end-to-end.

[0046] In summary, this application includes at least one of the following beneficial technical effects:

[0047] 1. By employing HVI color space transformation specifically designed for low-light scenes, Laplacian pyramid multi-scale decomposition, and differentiated feature extraction networks tailored to different scale characteristics, this method can more accurately characterize complex distortion patterns in low-light enhanced images. In particular, the complementary luminance and chromaticity information achieved through the Cross-Attention Interaction Module (CAIM) effectively models the coupled distortion relationship between the two. This significantly improves the accuracy and robustness of low-light enhanced image quality assessment.

[0048] 2. First, the luminance and chrominance information is separated from the source using the HVI space. Then, the CAIM (Cross-Attention Interaction Module) is introduced in the core stage of feature extraction, enabling bidirectional and adaptive information compensation and correction of luminance and chrominance features. This mechanism allows the network to understand prior associations such as low color confidence in overexposed areas, thereby achieving collaborative evaluation of luminance-chrominance coupling distortion.

[0049] 3. By employing a multi-scale decomposition and divide-and-conquer strategy, the Transformer module is used to specifically capture global illumination and structural information, while the enhanced residual module focuses on local details and texture distortion. This is then systematically integrated through a Progressive Feature Fusion (PFFM) module. This design not only makes the model conform to the multi-scale perception characteristics of human vision, achieving efficient coordination between global perception and local evaluation, but also optimizes computational efficiency through subsampling and other design features, balancing model performance and complexity. Attached Figure Description

[0050] Figure 1 This application presents examples of reconstructed images obtained after processing the same low-light scene with different enhancement algorithms and their corresponding average subjective opinion scores.

[0051] Figure 2 This is a schematic diagram of the framework of this application.

[0052] Figure 3 This is a flowchart illustrating the process of this application.

[0053] Figure 4 This is a flowchart illustrating the hierarchical feature extraction stage of this application.

[0054] Figure 5 This is a schematic diagram of the RASPP module framework in this application.

[0055] Figure 6 This is a schematic diagram of the framework of the Transformer module in this application.

[0056] Figure 7 This is a schematic diagram of the framework of the enhanced residual module in this application.

[0057] Figure 8 This is a schematic diagram of the HCAM module framework of this application.

[0058] Figure 9 This is a schematic diagram of the framework of the progressive feature interaction module of this application.

[0059] Figure 10 This is a scatter plot of the method of this application on three databases: LIEQ, LEISD, and SQUARE-LOL.

[0060] Figure 11 This is a t-test significance matrix plot of the method in this application on the LIEQ, LEISD, and SQUARE-LOL databases. Detailed Implementation

[0061] The following is in conjunction with the appendix Figure 1-11 This application will be described in further detail.

[0062] Image quality assessment methods without reference aim to directly evaluate the perceived quality of an image without relying on a reference image. Traditional methods mainly follow two technical routes: statistical modeling methods based on handcrafted features and feature self-learning methods based on deep learning. Statistical modeling methods based on handcrafted features rely on manually designed statistical features of natural scenes. Early methods quantified distortion by extracting statistical features of coefficients in the transform domain (such as the DCT domain). Subsequent research operated directly in the spatial domain, measuring the "naturalness" of an image by calculating the statistical properties of locally normalized brightness coefficients. Other methods predict quality by measuring the deviation between the image and a pre-built statistical model, or by evaluating the correlation between global statistics such as image entropy and distortion type. To improve the ability to distinguish different distortions, some methods attempt to apply additional distortion to the image to enhance feature recognition. Recent research has strengthened the perception of key channels such as brightness by introducing attention mechanisms, or combined multi-scale transform domain features for quality assessment. Although these handcrafted feature methods have good interpretability, their feature representation capabilities are limited, making it difficult to characterize the complex, coupled nonlinear distortions in low-light enhanced images.

[0063] With the development of deep learning, researchers have begun to utilize deep neural networks to learn the mapping relationship from images to quality scores end-to-end. Early work used multi-layer convolutional and pooling structures to build the basic evaluation framework. To improve performance, subsequent methods introduced dual-branch network structures, fusing features from different structures or tasks through mechanisms such as bilinear pooling. Further research focused on more refined feature interactions and fusions, such as fusing local and global features through adaptive mechanisms, or recovering degradation information to aid evaluation through visual compensation and reconstruction strategies. In recent years, the Transformer structure has been introduced into this field to extract global image features and combine them with ranking loss for fine-grained quality prediction. Meanwhile, some studies have focused on jointly modeling local and global feature interactions and incorporating multi-scale attention mechanisms to capture structural and texture degradation. Furthermore, predicting error maps or pre-training on large-scale synthetic data to obtain more discriminative feature representations has also become a way to improve model performance.

[0064] However, the aforementioned methods share common limitations, posing challenges for low-light enhanced image quality assessment. Firstly, whether based on handcrafted features or deep learning, existing methods are largely designed for general natural images, and their feature representations are not optimized for the specific distortion patterns arising from insufficient illumination, sensor noise, and the combined effects of enhancement algorithms in low-light environments. The distortion in low-light enhanced images is more complex, exhibiting severe coupled distortion of brightness and chromaticity information (e.g., overexposed areas are accompanied by color distortion, while underexposed areas suffer from noise and loss of detail).

[0065] like Figure 1As shown, this example displays reconstructed images obtained by different enhancement algorithms after processing the same low-light scene, along with their corresponding average subjective opinion scores. It can be observed that the subjective score results are closely related to the brightness and chromaticity performance of the reconstructed images. Brightness undercompensation (such as...) Figure 1 (as indicated by the arrow in (a1)) will cause the overall scene to be too dark, making it difficult to discern details in the target area; while over-compensation of brightness (such as...) will cause the scene to be too dark, making it difficult to distinguish details in the target area; and over-compensation of brightness (such as...) Figure 1 (As indicated by the arrow in (a5)) will cause overexposure and loss of detail in local areas. This shows that a moderate global and local brightness level is crucial for perceived quality. Furthermore, different enhancement algorithms show significant differences in color reproduction, and may even exhibit noticeable color shift distortion (such as...). Figure 1 (As indicated by the arrow in (b5), the sky area is incorrectly enhanced to an unnatural pink). Therefore, the quality degradation of low-light enhanced images is highly dependent on the accurate and coordinated restoration of luminance and chromaticity information in the image.

[0066] In summary, the visual quality defects of low-light enhanced images exhibit complex cross-modal and cross-scale characteristics, including "deep coupling of luminance and chromaticity distortion" and "degradation manifesting in both global illumination and local details." This means that designing a no-reference quality assessment model for such images must be able to decouple and interactively analyze luminance and chromaticity information while simultaneously considering the feature representation of both global structure and local details. Existing general assessment methods struggle to accurately capture these complex degradation characteristics, resulting in insufficient consistency between their predictions and human subjective perception in low-light enhanced scenes.

[0067] Based on this, the present invention proposes a referenceless quality assessment scheme specifically for low-light enhanced images. Figure 2 The overall framework of the proposed method is outlined, comprising three main stages: image transformation and decomposition, hierarchical feature extraction, and feature fusion and regression. Specifically, firstly, HVI color space transformation is used to decouple the image signal, resulting in the luminance component (I) and chrominance component (HV). Then, Laplacian pyramid decomposition is applied independently to both components to generate multi-scale representations. The lowest scale represents the low-frequency content of the image, while the remaining scales represent high-frequency details at different granularities. Based on this, a subsampling Transformer module is designed to extract global features from the low-frequency content, and an enhanced residual block is constructed to extract local features from the high-frequency details. Simultaneously, to fully exploit the complementarity between luminance and chrominance, a cross-attention interaction module is introduced to establish a connection between them. Hierarchical feature representation is achieved by introducing pooling-based downsampling and stacking multiple of the aforementioned feature extraction and interaction modules. Finally, a progressive feature fusion module is constructed to integrate deep multi-scale features, which are then regressed to obtain a quality score.

[0068] This scheme systematically addresses the core challenges of existing methods in assessing the quality of low-light enhanced images: the difficulty in separating luminance-chrominance coupled distortion and the inability to simultaneously address multi-scale degradation features. The following sections will elaborate on the various components and workflow of this scheme.

[0069] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0070] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article, unless otherwise specified, generally indicates that the preceding and following related objects have an "or" relationship.

[0071] The embodiments of the present invention will now be described in further detail with reference to the accompanying drawings.

[0072] Reference Figure 3 This invention provides a referenceless quality assessment method for low-light enhanced images, the main process of which is described below.

[0073] Step S100: Obtain the low-light enhancement image to be evaluated.

[0074] In practical applications, the input image can come from various edge IoT devices (such as surveillance cameras, vehicle cameras, drones, etc.) and be the output after processing by any existing low-light enhancement algorithm. Alternatively, it can be a low-light enhanced image obtained by computer devices or mobile devices after processing low-darkness images imported by the operator in various computer programs or apps.

[0075] Step S200: Convert the low-light enhanced image from the original color space to the HVI color space to separate the luminance component and chrominance component.

[0076] The original color space refers to the RGB space, which is generally used to represent low-light enhanced images. However, the three RGB channels are highly interdependent, making it difficult to accurately describe luminance and chromaticity information. Furthermore, even if noise disturbances only interfere with luminance information, their impact on all channels in the RGB space increases the difficulty of analysis. Therefore, this application adopts the HVI color space, which explicitly decouples luminance and chromaticity information and includes learnable parameters to adapt to different luminance conditions.

[0077] Specifically, the intensity channel of each pixel is first calculated to estimate the intensity / brightness map of the input image, as shown below:

[0078]

[0079] Where I represents the estimated intensity / luminance map, c is the RGB color channel index, and H and W are the height and width of the image.

[0080] Subsequently, to obtain color representations insensitive to changes in illumination, the RGB image was initially mapped to the HSV color space. In the HSV space, the value (V) component corresponds to the aforementioned intensity mapping (i.e., V = I), while the HS plane, composed of hue (H) and saturation (S), provides a chromaticity description basis relatively independent of illumination. The Hue axis, after polarization, yields orthogonal h and v, used to describe the horizontal and vertical components of color in the two-dimensional chromaticity plane. Simultaneously, considering that noise intensity in low-light images is often negatively correlated with luminance values ​​(i.e., noise is more significant in dark areas), directly using the original luminance value I to modulate chromaticity information would amplify the impact of dark noise. To adapt to the non-uniformity of noise in different brightness regions under low-light conditions, a learnable intensity collapse function Γ is introduced. α Its form is:

[0081]

[0082] Where α is a trainable parameter used to adjust the density of the dark areas, and ε is a small constant, typically ε = 1 × 10⁻⁶. -8 To avoid gradient explosion.

[0083] Finally, the intensity collapse function is used to adaptively weight the chromaticity information derived from the HSV space to generate noise-robust horizontal and vertical chromaticity maps. The chromaticity map can ultimately be expressed as a horizontal map. ) and vertical map ( )form:

[0084]

[0085] Where h∈H and v∈V, S is the saturation map obtained by the HS-plane transformation, and ⨀ denotes element-wise multiplication. Thus, the original image can be decomposed into three parts: luminance component... and horizontal components and vertical components The two together constitute the chromaticity component (HV). This conversion process not only achieves explicit separation of luminance and chromaticity but also enables the model to adaptively suppress the interference of dark noise on the chromaticity signal, laying a solid foundation for subsequent accurate quality assessment. Its visualization results are as follows: Figure 2 As shown on the left.

[0086] Step S300: Perform multi-scale decomposition on the luminance component and the chrominance component respectively to obtain sets of luminance sub-components and sets of chrominance sub-components corresponding to multiple different scales.

[0087] As mentioned earlier, after color space conversion, we obtained the separated luminance component I and chrominance component HV. To model the distortion at different spatial frequencies in the image, this embodiment performs multi-scale decomposition on these two components. Scale decomposition refers to breaking down each component into different scales; large-scale components typically correspond to low-frequency components, while small-scale components typically correspond to high-frequency components. The purpose of scale decomposition is to separate image information in both frequency and space. The large-scale (low-resolution) components mainly contain low-frequency information of the image, such as the overall illumination distribution, scene contrast, and contour structure, which are crucial for determining whether the image is overall overexposed or underexposed. The small-scale (high-resolution) components mainly contain high-frequency information of the image, such as object textures, edges, and artifacts or noise generated by enhancement algorithms, which relate to the naturalness and sharpness of local details. Through this explicit separation, subsequent differential analysis of distortions of different natures can be performed, thereby more comprehensively evaluating image quality.

[0088] The multi-scale decomposition employs Laplace pyramid decomposition. Specifically, let L∈{I,HV}, and first generate a Gaussian pyramid through Gaussian filtering and downsampling:

[0089]

[0090] in, The k-th scale image of the Gaussian pyramid is obtained by applying Gaussian filtering and downsampling to the previous scale. The kernel is a normalized Gaussian kernel; * denotes a two-dimensional discrete convolution. This represents a downsampling operation with a factor of 2. The purpose of constructing a Gaussian pyramid is to obtain a series of progressively lower resolution, increasingly blurry versions of the image. Lower resolution images preserve the original low-frequency information (subject outlines, large areas of light and dark), but lose high-frequency details (texture, edges). First, the original image is smoothed using a small Gaussian kernel (convolution operation*). The purpose of this is to prevent aliasing distortion during downsampling. The smoothed image is then halved in size (i.e., downsampling), and this process is repeated to generate a series of images.

[0091] Subsequently, the Laplace pyramid is constructed by differencing adjacent Gaussian layers:

[0092]

[0093] in, An upsampling factor of 2 is used to restore scale matching with the previous scale. The lowest resolution scale does not require differencing. The Laplacian pyramid is constructed to record high-frequency details lost during blurring and downsampling at each layer. Specifically, for a given layer of the Gaussian pyramid, it is first magnified back to the size of the previous layer (i.e., an upsampling operation). This magnification process is approximate and loses some information. Then, the difference between this magnified, blurred image and the sharper Gaussian image from the previous layer is calculated. This difference is the Laplacian pyramid image for that layer. It captures high-frequency details such as edges and textures that cannot be predicted by lower layers.

[0094] In this application, we set the scale to 3 to balance the granularity and complexity of the scale decomposition. Finally, we obtain the sets of luminance and chrominance subcomponents corresponding to the decomposition scale.

[0095] Step S400: At the same scale, perform hierarchical feature extraction and interactive processing on the corresponding luminance subcomponent and chrominance subcomponent respectively to obtain the enhanced luminance features and enhanced chrominance features at that scale after interactive enhancement.

[0096] Specifically, this step first utilizes convolutional layers to map multi-scale luminance / chrominance components to the feature space. Then, a subsampling Transformer module and an enhanced residual module are used to extract global and local features at different scales, respectively. A cross-attention interaction module is introduced to achieve feature interaction between luminance and chrominance components. By integrating max pooling and stacking multiple feature extraction and interaction modules, a hierarchical feature extraction process from shallow to deep is achieved. (Refer to...) Figure 4 Specifically, it includes:

[0097] Step S410: Perform shallow feature extraction on the luminance subcomponent and chrominance subcomponent at different scales to obtain shallow luminance features and shallow chrominance features.

[0098] This step aims to perform basic feature mapping and contextual information aggregation on each component obtained after multi-scale decomposition, providing richer feature representations for subsequent deep processing. For each luminance and chrominance sub-component in the Laplacian pyramid, a 3×3 convolution layer is first applied independently to map pixel information to the latent feature space. Then, a Residual Atrous Spatial Pyramid Pooling (RASP) module is used to extract shallow features for each scale component. The above process is represented as follows:

[0099]

[0100] in, This represents the shallow features obtained, corresponding to scale k. Represents the input luminance / chrominance components, Conv 3×3 (•) represents a 3×3 convolution layer, and RASPP(•) represents a RASPP module. Correspondingly, after processing the luma and chroma subcomponents at each scale, the shallow luma and chroma features at the corresponding scale can be obtained.

[0101] Figure 5 The structure of the RASPP module is presented, consisting of three parallel 3×3 convolutions with different dilation rates and one 1×1 convolution. Specifically, it includes three parallel 3×3 dilated convolution branches with dilation rates of 1, 2, and 4, respectively, to capture contextual information under different receptive fields. The outputs of all branches are concatenated along the channel dimension, and all feature information is aggregated and the feature channels are compressed through a 1×1 convolution. Finally, the shallow features are output by adding the residual connections to the module input.

[0102] Traditional methods often use pooling operations to expand the receptive field, but this irreversibly reduces the spatial resolution of the feature map, making it difficult to evaluate image detail distortion. The RASPP module, through dilated convolution, enables the network to obtain powerful multi-scale context awareness at shallow layers while maintaining the spatial size of the feature map, thus laying the foundation for accurate local and global quality assessment.

[0103] Step S420: Perform deep feature extraction on the shallow luminance features and shallow chrominance features respectively to obtain deep luminance features and deep chrominance features.

[0104] This step includes processing at least one low-frequency scale component using a self-attention-based module, and processing other scale components using a convolutional neural network-based module.

[0105] Given a scale of 3, the lowest-scale features obtained from the Laplacian pyramid decomposition primarily contain overall brightness / chromaticity information. Therefore, the subsampling Transformer module is used here to extract global features and model a long-range, cross-regional structure, thus obtaining the corresponding deep brightness and deep chromaticity features for this component. Specifically, the input to this step is the shallow brightness and shallow chromaticity features obtained from shallow feature extraction in S410.

[0106] like Figure 6 As shown, the proposed SS-Trans module mainly consists of two parts: Self-Attention (SA) and Feed-Forward Network (FFN), which are used to explicitly establish global relevance and enhance feature representation capabilities, respectively. Specifically, for the input features after layer normalization, a single 1×1 convolution is used to project them as the query Q, while two parallel 2×2 convolutions with a stride of 2 are used to downsample them, and then a 1×1 convolution is used to project them as the key K and the value V, respectively, as shown below:

[0107]

[0108] Where R(•) represents the Reshape operation, Conv1×1(•) represents a 1×1 convolution, and Conv2×2(•) represents a 2×2 convolution. It should be noted that due to the increase in network depth, the feature resolution decreases, so 2×2 convolutions are only applied to the first two feature levels to reduce computational complexity.

[0109] Next, matrix multiplication is used to compute multi-head self-attention between Q and K to describe the global dependencies between spatial locations, and the result is multiplied with V to obtain the attention-weighted feature F. A As shown below:

[0110]

[0111] Where Softmax(•) represents the Softmax activation function, D is a scaling factor, and the superscript T indicates transpose.

[0112] A single 1×1 convolutional layer is used to project the feature expression (FA) onto the input features, and the result is element-wise added to obtain the self-attention output features. Since the above process primarily considers global dependencies and loses local patterns, a simple feedforward network is introduced to enhance the local representation of features and implicitly embed positional encodings. For example... Figure 6 As shown on the right, the extracted The data is further fed into a feedforward network (FFN), which contains LayerNorm, 1×1 convolutions, 3×3 depthwise separable convolutions, GELU activation, and a final 1×1 convolution to enhance the local correlation and non-linear representation of features. Finally, the output of the FFN is... Element-wise addition is performed, and the final feature is obtained through the residual mechanism. This enables stable and efficient global modeling.

[0113] When the scale is greater than 3, at least one shallow luminance feature and shallow chrominance feature containing the lowest scale can be used to extract the deep luminance feature and deep metric feature of the corresponding scale through the subsampling Transformer module. For example, when the scale is 4, the lowest scale and the second lowest scale can be processed accordingly.

[0114] For the large-scale components obtained from the Laplacian pyramid decomposition, which contain texture details at different granularities and exhibit strong local correlations, an enhanced residual module with a local receptive field is used for deep feature extraction. Here, the large-scale components are the shallow luminance and chrominance features of other layers that were not extracted by the subsampling Transformer module.

[0115] Figure 7 The structure of the proposed augmented residual block is presented, the core of which lies in introducing parallel spatial and channel attention to focus on rich features while ignoring useless ones. Specifically, the input features are first passed through two 3×3 convolutional layers to extract local features. These features are then fed into the attention block for feature importance weighting. Spatial attention uses a 1×1 convolution to compress the channel dimension of the features, followed by a 7×7 convolution to capture contextual information and generate spatial weights activated by a sigmoid function. These spatial weights guide the network to focus on key spatial locations rich in detail. Channel attention performs global average pooling on the features to obtain channel information, then uses two 1×1 convolutions to learn channel correlations, and outputs channel weights through a sigmoid function. These channel weights adaptively adjust the importance of each feature channel. Finally, the two attention weights are element-wise multiplied with the previously extracted features to achieve feature weighting, and a residual connection is integrated to generate the final spatially and channel-enhanced features.

[0116] After this step, the shallow luminance features and shallow chrominance features at all scales will be extracted into deep luminance features and deep chrominance features respectively.

[0117] Step S430: Interact with deep luminance features and deep chrominance features at the same scale through a cross-attention mechanism to obtain enhanced luminance features and enhanced chrominance features.

[0118] Although illuminance and chromaticity physically represent light intensity and color respectively, they are coupled during image degradation. To achieve feature complementarity between the two in the deep feature extraction stage, a CAIM (Cross-Attention Interaction Module) is introduced after each feature extraction. Specifically, deep illuminance features are used as query vectors, and deep chromaticity features are used as key and value vectors. A first cross-attention is calculated, and its result is used to enhance the deep illuminance features to obtain enhanced illuminance features. Similarly, deep chromaticity features are used as query vectors, and deep illuminance features are used as key and value vectors. A second cross-attention is calculated, and its result is used to enhance the deep chromaticity features to obtain enhanced chromaticity features.

[0119] like Figure 8 As shown, the proposed CAIM contains two branches with the same structure: one uses luminance features to compensate for chrominance features, and the other does the opposite, which can be represented as follows:

[0120]

[0121] Among them, F I / F HV This represents the deep luminance / chrominance features of the input. / Indicates the enhanced luminance / chrominance characteristics of the output, Inter I-HV (•) and Inter HV-I (•) represents the chroma compensation luminance branch and the luminance compensation chroma branch, respectively.

[0122] For the sake of brevity, we will use the left branch (chroma enhancement for luminance) as an example. The right branch is the exact opposite of the left branch, and will not be elaborated further. Specifically, for the deep luminance features after layer normalization, we first use cascaded 1×1 convolutions and 3×3 depthwise convolutions to encode pixel-level channel context and channel-level spatial context to generate the query vector Q. I Simultaneously, the same convolutional structure is used to project deep chromaticity features to generate the key vector K. HV Sum vector V HV Then, Q is calculated using matrix multiplication followed by Softmax activation.I and K HV Cross-attention is used to establish long-range correlation between the two. Based on this, cross-attention is combined with V. HV The compensated chromaticity features are obtained by matrix multiplication and then projected onto the original luminance features through a 1×1 convolution.

[0123] Finally, a gated feedforward network is used to further enhance the brightness features, such as... Figure 8 As shown in the bottom left. In short, 1×1 convolutions are used to increase feature dimensionality, and two parallel 3×3 depthwise convolutions (one activated by GeLU) are used to model local patterns and learn gated attention, respectively. Finally, a single 1×1 convolution and residual connections are used to output enhanced brightness features.

[0124] In the same way, enhanced chromaticity features can be obtained, such as Figure 8 As shown in the lower right corner. It should be noted that this CAIM is applied to the interaction of luminance and chrominance features at each scale, thereby promoting subsequent cross-scale feature fusion.

[0125] Step S500: Fuse the enhanced luminance features and enhanced chrominance features from all scales to generate the final fused feature.

[0126] Reference Figure 9 To fully integrate luminance and chrominance features at different scales, this application employs a Progressive Feature Fusion Module (PFFM) for integration. First, at each scale, fusion weights are adaptively generated based on the enhanced luminance and chrominance features for that scale. Specifically, for a given scale k∈{0,1,2}, the luminance and chrominance features are concatenated at each scale, and the concatenated result is fed into a 1×1 convolutional layer to generate 2-channel fusion weights, which are then normalized using the Softmax function, as shown below:

[0127]

[0128] in, [•] indicates the generated fusion weight, and [•] indicates channel cascading.

[0129] Subsequently, the enhanced luminance features and enhanced chrominance features at this scale are weighted and fused according to the fusion weights to obtain the fused features at this scale, as shown below:

[0130]

[0131] in ⊙ indicates fusion feature, and ⊙ indicates element-level multiplication.

[0132] Since the resolution of the fused features differs across scales, the fused features at different scales are then aligned for resolution. Following a descending order from high-resolution to low-resolution scales, the high-resolution features are progressively fused into the low-resolution features. Specifically, the fused features at the largest scale (k=2) undergo two 2×2 max pooling operations, the fused features at the second largest scale (k=1) undergo one 2×2 pooling operation, and the fused features at the smallest scale (k=0) are left unprocessed, thus achieving cross-scale resolution alignment. When the scale decomposition is greater than 3, the number of pooling operations is increased accordingly. When the scale is 3, this step is specifically represented as follows:

[0133]

[0134] Pool 2×2 (•) indicates max pooling.

[0135] Based on the above, and considering the complementarity of features at different scales, element-wise addition is used to achieve cross-scale feature interaction. Simply put, large-scale features are added to small-scale features scale by scale to compensate for their local details. Then, the previously mentioned residual enhancement module is used to refine the interacting multi-scale features, as shown below:

[0136]

[0137] Here, E-Res(•) represents the aforementioned residual enhancement module.

[0138] Finally, the refined features from different scales are concatenated along the channel dimension, and a 1×1 convolution is used to achieve cross-scale compression and integration, resulting in the final fused features, as shown below:

[0139]

[0140] Step S600: Based on the final fusion features, regress to obtain the quality assessment score of the low-light enhanced image.

[0141] In the regression phase, the final fused features are subjected to global average pooling to obtain a one-dimensional feature vector. This vector discards all spatial location information but retains the global strength of the quality attribute represented by each channel. This forces the network to learn a spatially independent, holistic quality representation, enhancing the model's robustness. The resulting one-dimensional feature vector is then input into a multilayer perceptron for nonlinear mapping to output the quality assessment score Q. Specifically:

[0142]

[0143] Where GAP(•) represents global mean pooling, and MLP(•) represents multilayer perceptron.

[0144] To fully verify the effectiveness of the proposed method, this application will describe a comparison of the results obtained after the method has been tested.

[0145] First, to comprehensively compare the proposed method with other comparative methods in terms of perceptual degradation modeling capabilities in low-light enhancement scenes, this application conducted a series of experiments on three publicly available datasets: LIEQ, LEIS, and SQUARE-LOL. Specifically, the LIEQ dataset contains 1000 enhanced images, generated from 100 original low-light images using 10 typical low-light enhancement algorithms; the LEIS dataset contains 2040 enhanced images covering 255 scenes, composed of 8 enhancement results; and the SQUARE-LOL dataset contains 2900 enhanced images covering 290 scenes, also constructed from 10 enhancement results.

[0146] Secondly, three mainstream quality evaluation metrics were used to assess model performance: Pearson linear correlation coefficient (PLCC), Spearman rank correlation coefficient (SROCC), and root mean square error (RMSE). SROCC primarily measures the monotonic consistency between the predicted results and subjective scores, while PLCC and RMSE focus on reflecting the accuracy and error level of the predicted scores. Higher SROCC and PLCC values, and lower RMSE values, indicate a greater consistency between the model's predicted results and the actual visual quality, resulting in better overall performance.

[0147] The experiments were conducted using the PyTorch deep learning framework. The experimental environment consisted of an AMD Ryzen 97900X 12-Core CPU (4.70 GHz), 64 GB of RAM, and an NVIDIA GeForce RTX 4090 (24GB). For network training, the Adam optimizer with β1=0.9 and β2=0.999 was used, with an initial learning rate of 2×10⁻⁶. -4The proposed network decays at a scaling factor of 0.5 every 100 epochs, with a batch size of 16. It is trained for 300 epochs to achieve convergence. During training, the dataset is divided according to scene content, with 80% of the scenes used for training and the remaining 20% ​​for testing. For data augmentation, the input image is randomly cropped into 224×224 image patches, and horizontal and vertical flipping is performed. During testing, to avoid randomness, 14 patches are randomly cropped from the input image, and the average of the 14 predicted quality scores is taken as the final score for that input image. This training-test process is repeated 10 times, and the median of the 10 results is used as the model score.

[0148] To fully validate the effectiveness of the proposed method, we conducted comparative experiments with three main types of BIQA methods: 1) Natural image BIQA methods based on handcrafted features, including BRISQUE, PSQA, and CTQA. These methods primarily rely on handcrafted features for quality prediction and offer strong interpretability. 2) Natural image IQA methods based on deep learning, including DBCNN, HyperIQA, VCRNet, TReS, VIPN, and TempQT. These methods learn image quality mapping relationships through networks, resulting in stronger feature representation capabilities. 3) BIQA methods for augmented scenes, including contrast-enhanced BIQA (NIQMC and BIQME) and low-light-enhanced BIQA (NLIEE, MFRQA, IACA, and JLCLNet). These methods focus more on the perceived quality after enhancement.

[0149] Table 1 presents the performance of the proposed method and the comparison methods on three publicly available low-light enhancement quality databases. The experimental results show that: 1) Traditional natural image BIQA methods based on handcrafted features generally perform poorly, especially on the LIEQ and LEISD datasets, where PLCC and SROCC are generally below 0.75. This is mainly because handcrafted statistical features cannot fully characterize the complex degradation factors present in low-light enhancement, thus limiting prediction accuracy. 2) Deep learning-based BIQA methods show a significant overall improvement, indicating that they can more effectively model the mapping from image features to visual perception. However, the training data for these models are of traditional distortion types, lacking adaptation to the degradation characteristics of low-light enhancement, thus still exhibiting performance bottlenecks on datasets such as LEISD where enhancement intensity is uneven or over-enhanced is more pronounced. 3) Contrast enhancement BIQA methods perform the weakest, indicating domain bias in low-light enhancement scenes. Low-light enhancement degradation involves not only contrast changes but also complex distortions such as color shift and overexposure, which traditional contrast enhancement BIQA metrics cannot effectively represent, resulting in low correlation between predictions and subjective ratings. 4) The BIQA method, specifically designed for low-light enhancement scenes, has a higher performance ceiling; for example, only IACA and JLCLNet achieve a SROCC exceeding 0.88 on SQUARE-LOL. This validates the necessity of targeted modeling strategies (such as structural, luminance, and chromaticity features) in this task. The proposed method achieves state-of-the-art or near-state-of-the-art performance on three databases, with PLCCs of 0.891, 0.831, and 0.911, respectively. This indicates that the proposed method can more fully capture the characteristics of low-light enhancement quality variations and possesses strong cross-scene robustness and generalization ability.

[0150] Table 1

[0151] Figure 10 The scatter plots of the proposed method on the LIEQ, LEISD, and SQUARE-LOL databases are shown. The horizontal axis represents the predicted score, and the vertical axis represents the MOS (Mean Orientation of Spectrum). Each point corresponds to a test sample. The diagonal line in the plot indicates that the predicted score is completely consistent with the subjective rating; the closer the point is to the diagonal line, the more accurate the prediction. As can be seen from the plots, the scatter plots of the proposed method are relatively concentrated across the three databases, with most points closely surrounding the diagonal line. This indicates that the model can stably and accurately capture the overall quality changes of low-light enhanced images. Furthermore, no significantly deviated samples appear in the scatter plots, indicating that the proposed method is also robust when dealing with extreme or moderately difficult samples, further validating its generalization ability to different enhancement algorithms.

[0152] Figure 11The t-test significance matrices of the proposed method on three databases are presented, with a confidence level set at 95%. Rows in the matrix represent the benchmark method, and columns represent the comparison methods. In the matrix, 1 indicates that the row method significantly outperforms the column method; -1 indicates that the row method significantly underperforms the column method; and 0 indicates that there is no significant difference between the row and column methods. The results show that the diagonal lines are usually 0 because the method shows no difference compared to itself. Observing the row containing the proposed method, most cells are 1, indicating that the proposed method significantly outperforms other comparison methods in most cases. The absence of -1 or 0 indicates that it has stable and consistent performance under different datasets and augmentation algorithm conditions, further verifying the robustness and generalization ability of the model.

[0153] Table 2 shows the performance of the proposed method on the LIEQ and LEISD datasets in different color spaces, including YCbCr, Lab, HSV, and the HVI color space ultimately adopted in this paper. On the LIEQ dataset, the PLCC using YCbCr or HSV are 0.872 and 0.844, respectively, with Lab being slightly lower (PLCC 0.825). After adopting HVI, the PLCC improves to 0.891, showing higher prediction accuracy and robustness. A similar trend is observed on the LEISD dataset: while the performance of traditional color spaces is good, it is slightly lower than that of HVI, with HVI achieving the best PLCC at 0.831. The main reason for this performance difference may be that traditional color spaces are prone to generating noise artifacts in low-light enhanced images, leading to color discontinuities and thus limiting the prediction accuracy of subsequent models. In contrast, the HVI color space captures light intensity changes through the luminance component I, and the horizontal H and vertical V components encode the spatial distribution characteristics of color, achieving reasonable decoupling and fusion of luminance and color, thereby significantly improving the model's prediction performance across various metrics.

[0154]

[0155] Table 2

[0156] Table 3 shows the performance of the proposed method on the LIEQ and LEISD datasets at different Laplacian decomposition scales. The experimental results show that as the decomposition scale increases from 1 to 3, the PLCC and SROCC indices steadily improve, reaching optimal performance on both datasets (PLC 0.891 on LIEQ; PLCC 0.831 on LEISD). When the decomposition scale is further increased to 4 or 5, the performance slightly decreases or remains flat, indicating that excessive decomposition scales do not provide additional improvement. It is evident that an appropriate Laplacian decomposition scale can effectively separate different frequency information in an image, enabling the model to simultaneously capture the low-frequency overall brightness structure and high-frequency local detail texture, thereby improving the accuracy of quality prediction. However, excessive decomposition scales may introduce information redundancy or noise accumulation, making it difficult for the model to fully utilize features at each scale. Therefore, this application ultimately selects scale 3 as the optimal configuration to achieve a balance between capturing multi-scale features and maintaining information effectiveness.

[0157]

[0158] Table 3

[0159] Table 4 shows the ablation experiment results of the proposed model on the LIEQ and LEISD datasets under different module configurations. Scheme 1: w / o HVI, removing HVI color space conversion, i.e., using the original RGB image as input; Scheme 2: w / o Laplacian decomposition, removing Laplacian decomposition; Scheme 3: enhancing the low-frequency branch after Laplacian decomposition with residual blocks; Scheme 4: w / o interaction module, removing the interaction module; Scheme 5: w / o fusion module, removing the fusion module; Scheme 6: w / o attention (residual block), extracting channel and spatial attention from the enhanced residual block. It can be seen that in terms of feature representation, when RGB is used instead of HVI color space (#1), the overall performance decreases significantly (e.g., PLCC on LIEQ drops from 0.891 to 0.860), indicating that the decoupling of the chromaticity structure and luminance of HVI makes a key contribution to the representation of low light quality features. When the Laplacian multiscale decomposition (#2) is removed, i.e. only single-scale features are retained, the performance further declines, indicating that a reasonable multiscale decomposition can enrich the frequency domain information representation, thereby improving the quality assessment capability.

[0160] Regarding model structure, scheme #3, which replaces the low-frequency branch with a traditional CNN, shows a slight performance decrease but still outperforms most other configurations. This indicates that the low-frequency channel is primarily responsible for modeling global information, while the Transformer excels in cross-region dependency modeling, thus leading to performance improvement. Furthermore, removing the feature interaction module (#4) or fusion module (#5) both result in a significant performance drop (e.g., PLCC drops to 0.874 and 0.864 on LIEQ, respectively), validating the indispensable role of cross-scale information interaction and fusion in enhancing discriminative capabilities. Additionally, removing the attention structure (#6) from the enhanced residual block also impacts performance; for example, SROCC drops from 0.786 to 0.759 on LEISD, demonstrating that attention helps highlight key regions and suppress noise interference. Ultimately, the complete model achieves the highest performance on both datasets, indicating that the proposed HVI color construction, multi-scale Laplacian decomposition, cross-branch interactive fusion, and attention enhancement mechanisms possess cumulative gain and synergistic effects in low-light image quality prediction.

[0161]

[0162] Table 4

[0163] Table 5 shows the generalization performance of the proposed method on the real nighttime enhancement database EHND. It should be noted that EHND differs from the three synthetic nighttime enhancement databases used previously; it is collected from real low-light scenes and generated by various enhancement algorithms, thus presenting a greater challenge in achieving realism in terms of distortion. On the EHND database, the proposed method achieves relatively best performance (highest PLCC and SROCC, second lowest RMSE), indicating that the constructed model also possesses high sensitivity and reliability in perceiving and discriminating the enhancement quality in real-world nighttime conditions.

[0164]

[0165] Table 5

[0166] Table 6 lists the computational complexity of various deep learning low-light enhancement IQA methods at an input resolution of 224×224, including the number of model parameters and floating-point operations (FLOPs). All methods were tested under the same hardware and software environment to ensure fairness in the computational complexity comparison. As can be seen from the table, the number of parameters for the proposed method is only 7.11M, significantly smaller than HyperIQA (27.38M) and VIPN (46.42M), and comparable to JLCLNet (7.19M), indicating that the proposed method has good lightweightness in terms of model size. Regarding FLOPs, the proposed method has 7.91G, slightly higher than TReS (8.39G) and JLCLNet (2.27G), but much lower than VCRNet (10.27G), indicating that while maintaining high performance, the computational cost remains within an acceptable range. Considering both the number of parameters and FLOPs, the method in this application achieves a good balance between model complexity and computational efficiency, making it suitable for deployment in environments with limited hardware resources.

[0167] method Params (M) FLOPs (G) DBCNN 15.31 2.06 HyperIQA 27.38 4.34 VCRNet 16.66 10.27 TReS 152.45 8.39 VIPN 46.42 4.12 JLCLNet 7.19 2.27 Proposed 7.11 7.91

[0168] Table 6

[0169] Table 7 shows the prediction performance of the proposed method on the LIEQ and LEISD databases under different training-to-test ratios. As the training data ratio increases from 20% to 80%, PLCC, SROCC, and RMSE on both databases show a continuous improvement trend, indicating that more training samples enable the model to learn more stable and discriminative quality representations. Notably, even with a training ratio of only 20%, the model achieves relatively considerable performance, demonstrating good data utilization efficiency. Under the 80%-20% training-to-test split, the model achieves the highest PLCC and SROCC, and the lowest RMSE on both databases, demonstrating the most outstanding performance. Overall, the results show that the proposed method not only fully benefits from additional training data but also exhibits robust generalization ability under different training-to-test splits.

[0170]

[0171] Table 7

[0172] Based on the same inventive concept, embodiments of this application also disclose a referenceless quality assessment system for low-light enhanced images, comprising:

[0173] The image acquisition module is configured to acquire the low-light enhanced image to be evaluated.

[0174] The color space conversion module is configured to convert the low-light enhanced image from the original color space to the HVI color space to separate the luminance component and the chrominance component.

[0175] The multi-scale decomposition module is configured to perform multi-scale decomposition on the luminance component and the chrominance component respectively, to obtain a set of luminance sub-components and a set of chrominance sub-components corresponding to multiple different scales.

[0176] The hierarchical feature extraction and interaction module is configured to perform hierarchical feature extraction and interaction processing on the corresponding luminance sub-component and chrominance sub-component at the same scale, respectively, to obtain enhanced luminance features and enhanced chrominance features at that scale after interaction enhancement.

[0177] The feature fusion module is configured to fuse enhanced luminance features and enhanced chrominance features from all scales to generate the final fused feature.

[0178] The quality regression module is configured to regress a quality assessment score of the low-light enhanced image based on the final fusion features.

[0179] The hierarchical feature extraction and interaction module includes a shallow feature extraction submodule, a deep feature extraction submodule, and a cross-attention submodule. The shallow feature extraction submodule is configured to perform shallow feature extraction on the luminance and chrominance subcomponents at different scales to obtain shallow luminance features and shallow chrominance features. This shallow feature extraction includes using convolutional operations with different receptive fields to aggregate multi-scale contextual information. The deep feature extraction submodule is configured to perform deep feature extraction on the shallow luminance and shallow chrominance features respectively to obtain deep luminance features and deep chrominance features. The cross-attention submodule is configured to interact with the deep luminance and deep chrominance features at the same scale through a cross-attention mechanism to obtain enhanced luminance and enhanced chrominance features. The deep feature extraction submodule includes a subsampling Transformer unit and an enhancement residual unit.

[0180] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional modules is merely an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. The specific working process of the system, modules, and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0181] In the several embodiments provided in this application, it should be understood that the disclosed systems, modules, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0182] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0183] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0184] This invention also provides a method for training a no-reference quality assessment model for low-light enhanced images, comprising:

[0185] Step S1: Obtain a training dataset containing multiple low-light enhanced images and their corresponding subjective quality scores.

[0186] Step S2: Construct a quality assessment model, which is configured to perform the aforementioned no-reference quality assessment method.

[0187] Step S3: Using the low-light enhanced image as input, the subjective quality score as the supervision label, and the mean absolute error as the loss function, perform end-to-end training on the model.

[0188] Specifically, to narrow the gap between the model's predicted quality score Q and the corresponding subjective quality score MOS, this application uses Mean Absolute Error (MAE) as the loss function for training, to directly measure the degree of deviation of the predicted result from the subjective rating, as shown in the following formula:

[0189]

[0190] Among them, MOS iLet represent the subjective quality score of the i-th sample, Qi represent the predicted quality score of the model, and N be the total number of training samples.

[0191] This invention also provides a computer-readable storage medium storing data that can be loaded and executed by a processor to implement the steps described in the foregoing process.

[0192] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0193] The above description of the embodiments is only used to provide a detailed introduction to the technical solutions of this application. However, the description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention, and should not be construed as a limitation of the present invention. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this invention should be covered within the protection scope of this invention.

Claims

1. A referenceless quality assessment method for low-light enhanced images, characterized in that, Includes the following steps: Acquire the low-light enhancement image to be evaluated; The low-light enhanced image is converted from the original color space to the HVI color space to separate the luminance and chrominance components; The luminance component and the chrominance component are decomposed into multiple scales to obtain sets of luminance sub-components and sets of chrominance sub-components corresponding to multiple different scales. At the same scale, hierarchical feature extraction and interactive processing are performed on the corresponding luminance sub-component and chrominance sub-component respectively to obtain enhanced luminance features and enhanced chrominance features at that scale after interactive enhancement; wherein, the hierarchical feature extraction and interactive processing includes using a self-attention-based module to process at least one low-frequency scale component, and using a convolutional neural network-based module to process other scale components. The enhanced luminance features and enhanced chrominance features from all scales are fused to generate the final fused feature; Based on the final fusion features, a quality assessment score for the low-light enhanced image is obtained through regression.

2. The referenceless quality assessment method for low-light enhanced images according to claim 1, characterized in that, The multi-scale decomposition is the Laplace pyramid decomposition.

3. The referenceless quality assessment method for low-light enhanced images according to claim 2, characterized in that, The number of decomposition scales of the Laplace pyramid is 3.

4. The referenceless quality assessment method for low-light enhanced images according to claim 1, characterized in that, The hierarchical feature extraction and interactive processing specifically include: Shallow feature extraction is performed on the luminance subcomponent and chrominance subcomponent at different scales to obtain shallow luminance features and shallow chrominance features. The shallow feature extraction includes using convolution operations with different receptive fields to aggregate multi-scale contextual information. Deep feature extraction is performed on the shallow luminance features and shallow chrominance features respectively to obtain deep luminance features and deep chrominance features; Deep luminance features and deep chrominance features at the same scale interact through a cross-attention mechanism to obtain enhanced luminance features and enhanced chrominance features.

5. The referenceless quality assessment method for low-light enhanced images according to claim 4, characterized in that, In the deep feature extraction step: For at least one low-frequency scale component that represents global low-frequency information obtained through multi-scale decomposition, feature extraction is performed using a unit based on the Transformer architecture. For other scale components that represent local high-frequency information obtained through multi-scale decomposition, feature extraction is performed using an enhanced residual unit based on a convolutional neural network and incorporating an attention mechanism.

6. The referenceless quality assessment method for low-light enhanced images according to claim 4, characterized in that, The interaction via the cross-attention mechanism specifically includes: Using deep luminance features as query vectors and deep chrominance features as key and value vectors, calculate the first cross attention and use the result to enhance the deep luminance features to obtain enhanced luminance features; Using deep chroma features as the query vector and deep luminance features as the key and value vectors, a second cross-attention is calculated, and the result is used to enhance the deep chroma features to obtain enhanced chroma features.

7. The referenceless quality assessment method for low-light enhanced images according to claim 1, characterized in that, The enhanced luminance and enhanced chrominance features from all scales are fused to generate the final fused features, including: At each scale, based on the enhanced brightness features and enhanced chromaticity features of that scale, a fusion weight is adaptively generated, and the enhanced brightness features and enhanced chromaticity features of that scale are weighted and fused according to the fusion weight to obtain the fused features of that scale. The fused features at different scales are aligned for resolution, and the high-resolution scale features are fused into the low-resolution scale features step by step in order from high-resolution scale to low-resolution scale. The fusion results are then refined to obtain the final fused features.

8. The referenceless quality assessment method for low-light enhanced images according to claim 1, characterized in that, The process of obtaining a quality assessment score through regression based on the final fusion features specifically includes: The final fused features are subjected to global average pooling to obtain a feature vector; The feature vector is input into a multilayer perceptron for nonlinear mapping, and the quality assessment score is output.

9. A referenceless quality assessment system for low-light enhanced images, characterized in that, include: The image acquisition module is configured to acquire the low-light enhanced image to be evaluated; The color space conversion module is configured to convert the low-light enhanced image from the original color space to the HVI color space to separate the luminance component and the chrominance component. The multi-scale decomposition module is configured to perform multi-scale decomposition on the luminance component and the chrominance component respectively, to obtain a set of luminance sub-components and a set of chrominance sub-components corresponding to multiple different scales. The hierarchical feature extraction and interaction module is configured to perform hierarchical feature extraction and interaction processing on the corresponding luminance sub-component and chrominance sub-component at the same scale, respectively, to obtain enhanced luminance features and enhanced chrominance features at that scale after interaction enhancement; wherein, the hierarchical feature extraction and interaction processing includes using a self-attention-based sub-module to process at least one low-frequency scale component, and using a convolutional neural network-based sub-module to process other scale components. The feature fusion module is configured to fuse enhanced luminance features and enhanced chrominance features from all scales to generate the final fused feature; The quality regression module is configured to regress a quality assessment score of the low-light enhanced image based on the final fusion features.

10. A method for training a no-reference quality assessment model for low-light enhanced images, characterized in that, include: Obtain a training dataset containing multiple low-light enhanced images and their corresponding subjective quality scores; A quality assessment model is constructed, the model being configured to perform the referenceless quality assessment method as described in any one of claims 1 to 8; Using the low-light enhanced image as input, the subjective quality score as the supervision label, and the mean absolute error as the loss function, the model is trained end-to-end.

Images