A method for no-reference underwater image quality assessment based on multi-color space joint modeling
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUILIN UNIV OF ELECTRONIC TECH
- Filing Date
- 2026-03-09
- Publication Date
- 2026-05-29
Smart Images
Figure CN122116107A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision and image processing technology, specifically to a method for evaluating the quality of underwater images without reference based on joint modeling of multiple color spaces, which belongs to the application technology of artificial intelligence and deep learning in underwater image quality evaluation. Background Technology
[0002] Underwater imaging technology is widely used in marine resource exploration, underwater robot navigation, marine ecological monitoring, and underwater target identification. However, due to the absorption and scattering of light by water, as well as the effects of low-light environments and suspended particulate matter, underwater images are prone to degradation during the imaging process, including color shift, contrast reduction, detail blurring, and fogging. These degradation features significantly reduce the visual quality of the images, affecting the accurate identification and understanding of underwater scenes and targets by human observers.
[0003] To improve underwater image quality, researchers have proposed numerous underwater image enhancement methods to improve contrast, sharpness, and color reproduction. However, as enhancement methods continue to evolve, how to objectively, reliably, and in accordance with human visual perception characteristics to evaluate the enhancement results has become a crucial issue in the field of underwater image processing. The accuracy of underwater image quality assessment methods directly affects the evaluation results of enhancement algorithm performance and the practical application effectiveness of underwater vision systems.
[0004] In the field of image quality assessment, existing methods are generally classified into three categories based on whether they rely on a reference image: full-reference, sub-reference, and no-reference. Since it is often difficult to obtain an ideal, distortion-free reference image in real underwater scenes, a no-reference assessment paradigm is more suitable for underwater image quality assessment. Most existing no-reference underwater image quality assessment methods are based on artificially designed features, such as contrast, color shift, and sharpness metrics. However, the degradation process of underwater images typically exhibits complex nonlinear characteristics, and different degradation factors are spatially non-uniformly distributed, making it difficult for assessment methods based on a small number of low-level features to comprehensively reflect the overall perceived quality of the image.
[0005] In recent years, deep learning methods have been introduced into underwater image quality assessment tasks to improve prediction performance by automatically learning image features. Although these methods have improved the accuracy of quality prediction to some extent, most methods are still based on a single color space for modeling, failing to effectively distinguish the differential effects of brightness degradation and chromaticity distortion in underwater images. Furthermore, in the process of multi-feature fusion, existing methods often employ relatively simple fusion strategies, making it difficult to fully utilize the complementary information between different perceptual features, resulting in insufficient stability and generalization ability in complex underwater environments.
[0006] Therefore, it is necessary to propose a referenceless underwater image quality assessment method that can jointly model multi-color space information, enhance the feature representation capability of key degraded regions, and improve the stability of quality prediction, so as to meet the practical application needs in complex underwater environments. Summary of the Invention
[0007] The purpose of this invention is to overcome the problems of existing underwater image quality assessment methods, such as their inability to comprehensively characterize brightness degradation and chromaticity distortion features and their insufficient adaptability to complex underwater environments. This invention provides a referenceless underwater image quality assessment method based on multi-color space joint modeling. This method constructs a multi-branch residual network structure and introduces an attention mechanism to achieve collaborative modeling of multi-dimensional perceptual features of underwater images, thereby improving the stability and accuracy of quality prediction.
[0008] To achieve the above objectives, the present invention adopts the following technical solution:
[0009] A method for quality assessment of referenceless underwater images based on multi-color space joint modeling includes the following steps:
[0010] Step 1: Construct an underwater image quality assessment model, which includes RGB feature extraction branches, luminance feature extraction branches, and chrominance feature extraction branches. In addition, residual structures and attention mechanisms are introduced into the model to enhance the ability to express quality degradation features.
[0011] Step 2: Obtain an underwater image database containing subjective quality scores, train the underwater image database, iteratively optimize the model parameters through a loss function and perform nonlinear mapping on the model output to alleviate the scale difference between the predicted score and the subjective score, and obtain the trained underwater image quality assessment model.
[0012] Step 3: Input the underwater image to be evaluated into the trained underwater image quality assessment model, and output the corresponding quality score result.
[0013] Furthermore, the specific process of step 1 is as follows:
[0014] The input underwater image is converted from the RGB color space to the YCbCr color space, and the luminance channel Y and the chrominance channel CbCr are separated.
[0015] An RGB feature extraction branch is constructed, and a deep convolutional structure based on residual network is used to perform multi-layer feature extraction on the original RGB image to obtain the semantic and structural features of the image.
[0016] A luminance feature extraction branch and a chrominance feature extraction branch are constructed respectively, wherein the luminance feature extraction branch takes the Y channel as input and the chrominance feature extraction branch takes the CbCr channel as input;
[0017] Both the luminance and chrominance branches adopt a shallow residual network structure, and the luminance degradation features and chrominance distortion features are extracted in layers through a stage-by-stage residual modeling method.
[0018] Channel attention and spatial attention mechanisms are introduced into the RGB feature extraction branch to enhance the network's ability to express key degraded regions and discriminative features;
[0019] At the end of each residual stage, luminance and chromaticity features are injected into the RGB backbone features for fusion, thereby achieving collaborative modeling of features in multiple color spaces.
[0020] Furthermore, the specific process of step 2 is as follows:
[0021] Acquire an underwater image database containing subjective quality ratings;
[0022] The underwater images in the database are subjected to uniform size adjustment and data normalization processing;
[0023] The database is divided into a training set and a validation set according to a preset ratio;
[0024] The underwater image quality assessment model is trained by inputting the training set into it, and the error between the predicted score and the subjective quality score is constrained by the L1 loss function.
[0025] The gradient backpropagation algorithm is used to iteratively optimize the model parameters until the model converges;
[0026] The model output is nonlinearly mapped using a five-parameter Logistic function to mitigate the scaling difference between predicted and subjective scores.
[0027] The trained underwater image quality assessment model is obtained.
[0028] Furthermore, the specific process of step 3 is as follows:
[0029] The underwater image to be evaluated is input into the trained underwater image quality assessment model. Multi-scale features are extracted through RGB branch, luminance branch and chrominance branch respectively. The features of each branch are jointly modeled through a multi-branch feature fusion module. The fused features are then subjected to global average pooling to obtain a global quality representation vector. Finally, the final image quality score is output through a fully connected regression layer to characterize the overall perceived quality of the underwater image.
[0030] The present invention has the following beneficial technical effects:
[0031] This invention achieves effective decoupling of luminance and chromaticity information through joint RGB and YCbCr multi-color space modeling, enhancing the ability to characterize complex underwater degradation phenomena. It combines a multi-branch residual network to achieve layer-by-layer modeling and fusion of multi-scale features, and introduces channel and spatial attention mechanisms to enhance sensitivity to key degradation regions. Simultaneously, it mitigates the scale difference between predicted and subjective scores through nonlinear mapping, improving the stability and consistency of quality prediction. This method requires no reference image, is suitable for image quality assessment tasks in real underwater scenarios, and has strong practical value and promising prospects for wider application. Attached Figure Description
[0032] To make the technical solution of the present invention clearer and more complete, the embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the accompanying drawings are only for illustrating the technical concept and implementation of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, various equivalent substitutions or formal modifications can be made based on the content shown in the drawings without departing from the technical spirit of the present invention, and all such modifications should fall within the scope of protection of the present invention.
[0033] Figure 1 This is a flowchart of the present invention;
[0034] Figure 2 This is a schematic diagram of the network structure of the multi-color space joint modeling model of the present invention;
[0035] Figure 3 This is a schematic diagram of the CBAM attention mechanism of the present invention;
[0036] Figure 4 This is a visual comparison chart of the detection performance of the method of the present invention and other mainstream target detection methods in this embodiment of the invention. Detailed Implementation
[0037] The present invention will be further described in detail below with reference to specific embodiments. It should be understood that the embodiments are only used to illustrate the technical solutions of the present invention, and not to limit the present invention. Those skilled in the art can make various modifications or substitutions to the present invention without departing from the technical concept and scope of protection of the claims, and all such modifications or substitutions should fall within the protection scope of the present invention.
[0038] refer to Figure 1 As shown, a method for quality assessment of referenceless underwater images based on multi-color space joint modeling includes the following steps:
[0039] Step 1: Construct an underwater image quality assessment model, which includes RGB feature extraction branches, luminance feature extraction branches, and chrominance feature extraction branches. In addition, residual structures and attention mechanisms are introduced into the model to enhance the ability to express quality degradation features.
[0040] Step 2: Obtain an underwater image database containing subjective quality scores, train the underwater image database, iteratively optimize the model parameters through a loss function and perform nonlinear mapping on the model output to alleviate the scale difference between the predicted score and the subjective score, and obtain the trained underwater image quality assessment model.
[0041] Step 3: Input the underwater image to be evaluated into the trained underwater image quality assessment model, and output the corresponding quality score result.
[0042] Furthermore, the specific process of step 1 is as follows:
[0043] The input underwater image is represented as: First, the underwater image is converted from the RGB color space to the YCbCr color space, resulting in a luminance channel and a chrominance channel: Y and CbCr. , ; Construct RGB feature extraction branches, luminance feature extraction branches, and chrominance feature extraction branches respectively. In the first... The stage characteristics of the RGB backbone network output in each stage are denoted as follows: The characteristic representation of the brightness branch output is as follows: The chroma branch outputs feature representations at the same stage. ;
[0044] At the end of each residual stage, the luminance and chrominance features are injected into the RGB backbone features. The fusion process is represented as follows: The fused features serve as input for the next stage, enabling layer-by-layer collaborative modeling of multi-color space features.
[0045] An attention enhancement residual module is introduced into the RGB backbone branch, and its channel attention weights are calculated as follows:
[0046]
[0047] Where F represents the input feature;
[0048] The channel-weighted output is:
[0049]
[0050] The spatial attention weights are calculated as follows:
[0051]
[0052] The final attention enhancement feature is:
[0053]
[0054] in , , This represents the Sigmoid activation function. This indicates element-wise multiplication and broadcasting along spatial dimensions.
[0055] Furthermore, the specific process of step 2 is as follows:
[0056] Obtain the dataset containing subjective quality ratings:
[0057]
[0058] in Indicates the first Zhang input image, Give it its corresponding true subjective quality score. This represents the total number of images in the dataset. All images are uniformly adjusted to 224 × 224 as network input. The database is divided into training and validation sets in an 8:2 ratio.
[0059] During the model training phase, global average pooling is performed on the fused features from the fourth stage:
[0060]
[0061] in Output features for the final stage;
[0062] The quality prediction value is output through a fully connected regression layer:
[0063]
[0064] The error between the predicted value and the subjective rating is constrained using the L1 loss function:
[0065]
[0066] To mitigate the scaling discrepancy between predicted scores and subjective ratings, a five-parameter Logistic nonlinear mapping is applied to the model output:
[0067]
[0068] Where m(s) represents the mapping quality score, s represents the model output, and θ1 ∼ θ5 are the mapping parameters. The trained underwater image quality assessment model is obtained by iteratively optimizing the model parameters.
[0069] Furthermore, the specific process of step 3 is as follows:
[0070] The underwater image to be evaluated is input into the trained underwater image quality assessment model. Multi-scale features are extracted through RGB, luminance, and chrominance branches. The final fused features are obtained through a stage-by-stage fusion mechanism. Finally, a predicted quality score is output through global pooling and regression layers.
[0071]
[0072] The predicted values are used to characterize the overall perceived quality of underwater images.
[0073] Furthermore, this invention compares and visualizes its implementation with existing mainstream underwater image quality assessment methods through specific embodiments, as detailed in the following experiments. Figure 4 See Table 1. Experimental results show that in complex underwater degradation scenarios and image samples with subtle quality differences, the multi-color space joint modeling underwater image quality assessment method proposed in this invention can more accurately reflect the true perceived quality of the image, and its prediction results have higher consistency with subjective evaluations. It achieves superior performance in both correlation and error indices, verifying the stability, robustness, and practical application value of this invention in complex underwater environments.
[0074] Table 1. Comparison of UWMCQR with other mainstream underwater image quality assessment methods on the LUIQD dataset.
[0075]
[0076] In summary, this invention proposes a referenceless underwater image quality assessment method based on multi-color space joint modeling. By constructing a multi-branch collaborative structure of RGB, luminance (Y), and chrominance (CbCr), it achieves effective decoupling and fusion representation of luminance degradation and chrominance distortion features in underwater images. Combining a residual learning framework with channel and spatial attention mechanisms enhances the model's sensitivity and discrimination ability to key degradation regions. Simultaneously, a nonlinear mapping mechanism improves the consistency between prediction results and subjective perception, thereby achieving more stable, accurate, and robust quality assessment results in complex underwater environments.
[0077] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. For those skilled in the art, any equivalent substitutions or modifications made to the embodiments without departing from the technical concept and scope of the claims should fall within the scope of protection of the present invention.
Claims
1. A method for quality assessment of referenceless underwater images based on multi-color space joint modeling, characterized in that, Includes the following steps: Step 1: Construct an underwater image quality assessment model, which includes RGB feature extraction branches, luminance feature extraction branches, and chrominance feature extraction branches. In addition, residual structures and attention mechanisms are introduced into the model to enhance the ability to express quality degradation features. Step 2: Obtain an underwater image database containing subjective quality scores, train the underwater image database, iteratively optimize the model parameters through a loss function and perform nonlinear mapping on the model output to alleviate the scale difference between the predicted score and the subjective score, and obtain the trained underwater image quality assessment model. Step 3: Input the underwater image to be evaluated into the trained underwater image quality assessment model, and output the corresponding quality score result.
2. The method for evaluating the quality of referenceless underwater images based on multi-color space joint modeling according to claim 1, characterized in that, In step 1, the model converts the input underwater image from the RGB color space to the YCbCr color space, and then constructs feature extraction branches for the luminance channel Y and the chrominance channel CbCr respectively.
3. The method for evaluating the quality of referenceless underwater images based on multi-color space joint modeling according to claim 1, characterized in that, The RGB feature extraction branch employs a deep convolutional network based on the residual network ResNet50, used to extract semantic and structural features of the image.
4. The method for evaluating the quality of referenceless underwater images based on multi-color space joint modeling according to claim 3, characterized in that, The improved deep convolutional network based on the residual network ResNet50 enhances the perception of degraded regions in underwater images by introducing the channel and spatial attention mechanism CBAM, thereby improving the ability to express the features of degraded regions in underwater images.
5. The method for evaluating the quality of referenceless underwater images based on multi-color space joint modeling according to claim 1, characterized in that, Both the luminance feature extraction branch and the chrominance feature extraction branch are shallow convolutional networks constructed using residual network structures.
6. The method for evaluating the quality of referenceless underwater images based on multi-color space joint modeling according to claim 5, characterized in that, The shallow convolutional network adopts the same hierarchical residual modeling method as the backbone network. It models the sharpness features layer by layer through four stage-by-stage residual stages. Each stage consists of two standard residual blocks connected in series, and the spatial resolution is downsampled and the channel dimension is adjusted at the end of the stage.
7. The method for evaluating the quality of referenceless underwater images based on multi-color space joint modeling according to claim 1, characterized in that, At the end of each residual stage of the RGB feature extraction branch, luminance feature extraction branch, and chrominance feature extraction branch, luminance features and chrominance features are injected into the RGB features for fusion to achieve collaborative modeling of multi-color space features.
8. The method for evaluating the quality of referenceless underwater images based on multi-color space joint modeling according to claim 1, characterized in that, The loss function mentioned in step 2 is the L1 loss function, which is used to constrain the error between the model's predicted quality score and the subjective quality score.
9. The method for evaluating the quality of referenceless underwater images based on multi-color space joint modeling according to claim 1, characterized in that, In step 2, a five-parameter Logistic function is used to perform a nonlinear mapping on the model output to alleviate the scaling difference between the predicted score and the subjective score.
10. The method for evaluating the quality of referenceless underwater images based on multi-color space joint modeling according to claim 1, characterized in that, In step 3, the fused features are subjected to global average pooling and a quality score is output through a fully connected layer.