A method for enhancing underwater degraded images
By combining deep learning with physical models, the problem of local overprocessing or underprocessing in water degradation image processing is solved, achieving more scientific image enhancement effects and speed improvement, with low hardware computing requirements.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHANGCHUN UNIV OF SCI & TECH
- Filing Date
- 2023-03-11
- Publication Date
- 2026-05-29
AI Technical Summary
Existing technologies suffer from local overprocessing or underprocessing in underwater degradation image processing, and the physical models lack universality and accuracy, failing to estimate parameters with high accuracy in all underwater scenarios.
A deep learning approach combined with a physical model was adopted. By determining the sample UIEB dataset, global light parameters and attenuation coefficients were calculated. Image enhancement was performed using convolutional kernels with parameter-corrected linear units and dilated convolutional kernels. The pixel value mean squared error loss function was used for training. The Adam optimizer was selected as the optimizer, with a learning rate of 0.001 and 1000 training epochs.
It achieves a more scientific and reasonable enhancement of water degradation images, improves processing speed and hardware computing performance, and the algorithm is a lightweight network with low computational requirements.
Smart Images
Figure CN116109517B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of image enhancement and deep learning technology, specifically a method for enhancing images of water degradation. Background Technology
[0002] Current methods for enhancing degraded underwater images mainly fall into two categories: traditional enhancement techniques and traditional physical models. Existing underwater processing algorithms based on enhancement techniques are adjusted and designed according to the characteristics of underwater images, and can improve the subjective visual effect of images to varying degrees. However, the underwater imaging environment is complex, and these methods do not analyze the mechanism of underwater image degradation, nor do they utilize the mapping information between images before and after degradation. Algorithms that directly enhance single or multiple features of the entire image often result in local overprocessing or underprocessing. Underwater degraded image processing algorithms based on physical models summarize the image degradation rules and, combined with prior information, indirectly restore clear images through parameter estimation. However, these algorithms are limited by the effectiveness and universality of the prior information used. Currently, there is no highly reliable prior information that allows underwater image processing algorithms to estimate the parameters required for restoration with high accuracy in all underwater scenarios. In addition, the underwater environment is complex and diverse, and the imaging environment is harsh and changeable. Even highly generalized physical models cannot cover all imaging-related factors, which means that there is still room for improvement in the universality and accuracy of such algorithms. Therefore, we propose a method for enhancing degraded underwater images to solve the above problems. Summary of the Invention
[0003] (a) Technical problems to be solved
[0004] To address the shortcomings of existing technologies, this invention provides a method for enhancing images of water degradation, thus solving the problems mentioned in the background section.
[0005] (II) Technical Solution
[0006] To achieve the above objectives, the present invention specifically adopts the following technical solution:
[0007] A method for enhancing images of water degradation, comprising two stages;
[0008] The first phase includes the following steps:
[0009] Step A: Determine the sample UIEB dataset;
[0010] Step B: Determine the physical model for scientific underwater image imaging. The mathematical expression of the physical model is: O c (x)=[I c (x)-L c ]·R c (x)+L c ;
[0011] Step C: Calculate global illumination parameters from the input image. The global illumination calculation module includes two 5×5 convolutional kernels, one global mean pooling layer, and two 1×1 convolutional kernels. Each convolutional kernel group contains three kernels. The activation function for the convolution operation is the Parametric Corrected Linear Unit (PReLU). The mathematical expression for PReLU is:
[0012] The second phase includes the following steps:
[0013] Step D: Calculate the attenuation coefficient by combining the input image with global light parameters. The attenuation coefficient calculation module connects the ambient light calculation with the input image.
[0014] Step E: Image enhancement is performed by combining the input image with global illumination and attenuation coefficients. Each input image is used as its corresponding reference image as the training target. The loss function is based on the mean squared error of pixel values to calculate the difference between input and output. The mathematical expression is:
[0015]
[0016] Furthermore, in the mathematical expression of step C, x is the value at each position on the feature map of the neural network, and a is a learnable parameter. The advantage of this activation function is that it can prevent the gradient from becoming 0 when x is negative, so that the corresponding parameters can continue to be updated.
[0017] Furthermore, the attenuation coefficient calculation module in step D includes three groups of 3×3 dilated convolution kernels, with dilation rates of 1 / 2 / 5 for the three groups of dilated convolution kernels, eight dilated convolution kernels in each group and one group of 3×3 ordinary convolution kernels, with a total of three ordinary convolution kernels.
[0018] Furthermore, in the mathematical expression of step E, H, W, and C represent the height, width, and number of channels of the image, respectively, and x, y, and z represent the coordinates of the image in three dimensions. R(I) represents the output of the algorithm in this paper, and R(I) represents the reference target in the training dataset.
[0019] Furthermore, in step E, the optimizer for the neural network parameters is the Adam optimizer, the batch size used in deep learning training is 1, the learning rate is 0.001, and the number of training rounds is 1000.
[0020] (III) Beneficial Effects
[0021] Compared with existing technologies, the present invention provides a method for enhancing images of water degradation, which has the following beneficial effects:
[0022] 1. The method proposed in this invention is more scientific and reasonable based on deep learning combined with physical models, and can more realistically enhance water degradation images compared with traditional methods.
[0023] 2. The algorithm of this invention has a significant speed advantage in enhancing water degradation images.
[0024] 3. The algorithm of this invention is a lightweight network with low requirements for hardware computing performance. Attached Figure Description
[0025] Figure 1 This is a flowchart of the water degradation image enhancement method based on deep learning combined with a physical model according to the present invention;
[0026] Figure 2 This is a flowchart of the global optical computation of the present invention;
[0027] Figure 3 This is a flowchart for calculating the attenuation coefficient of the present invention. Detailed Implementation
[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0029] Example
[0030] like Figure 1-3 As shown, an embodiment of the present invention proposes a method for enhancing water degradation images, comprising two stages;
[0031] The first phase includes the following steps:
[0032] Step A: Determine the sample UIEB dataset;
[0033] Step B: Determine the physical model for scientific underwater image imaging. The mathematical expression of the physical model is: O c (x)=[I c (x)-L c ]·R c (x)+L c ;
[0034] Step C: Calculate global illumination parameters from the input image. The global illumination calculation module includes two 5×5 convolutional kernels, one global mean pooling layer, and two 1×1 convolutional kernels. Each convolutional kernel group contains three kernels. The activation function for the convolution operation is the Parametric Corrected Linear Unit (PReLU). The mathematical expression for PReLU is:
[0035] The second phase includes the following steps:
[0036] Step D: Calculate the attenuation coefficient by combining the input image with global light parameters. The attenuation coefficient calculation module connects the ambient light calculation with the input image.
[0037] Step E: Image enhancement is performed by combining the input image with global illumination and attenuation coefficients. Each input image is used as its corresponding reference image as the training target. The loss function is based on the mean squared error of pixel values to calculate the difference between input and output. The mathematical expression is:
[0038]
[0039] In this embodiment, x in the mathematical expression in step C is the value of each position on the feature map of the neural network, and a is a learnable parameter. The advantage of this activation function is that it can prevent the gradient from becoming 0 when x is negative, so that the corresponding parameters can continue to be updated.
[0040] In this embodiment, the attenuation coefficient calculation module in step D includes three groups of 3×3 dilated convolution kernels with dilation rates of 1 / 2 / 5 respectively. Each group has eight dilated convolution kernels and one group of 3×3 ordinary convolution kernels, with a total of three ordinary convolution kernels.
[0041] In this embodiment, H, W, and C in the mathematical expression of step E represent the height, width, and number of channels of the image, respectively, and x, y, and z represent the coordinates of the image in three dimensions. R(I) represents the output of the algorithm in this paper, and R(I) represents the reference target in the training dataset.
[0042] In this embodiment, the Adam optimizer is selected for the neural network parameters in step E, the batch size used in deep learning training is 1, the learning rate is 0.001, and the number of training rounds is 1000.
[0043] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A method for enhancing images of water degradation, characterized in that: It includes two stages; The first phase includes the following steps: Step A: Determine the sample UIEB dataset; Step B: Determine the physical model for scientific underwater image imaging. The mathematical expression of the physical model is: O c (x)=[I c (x)-L c ]·R c (x)+L c ; Step C: Calculate global illumination parameters from the input image. The global illumination calculation module includes two 5×5 convolutional kernels, one global mean pooling layer, and two 1×1 convolutional kernels. Each convolutional kernel group contains three kernels. The activation function for the convolution operation is the Parametric Corrected Linear Unit (PReLU). The mathematical expression for PReLU is: The second phase includes the following steps: Step D: Calculate the attenuation coefficient by combining the input image with global light parameters. The attenuation coefficient calculation module connects the ambient light calculation with the input image. Step E: Image enhancement is performed by combining the input image with global illumination and attenuation coefficients. Each input image is used as its corresponding reference image as the training target. The loss function is based on the mean squared error of pixel values to calculate the difference between input and output. The mathematical expression is:
2. The method for enhancing water degradation images according to claim 1, characterized in that: In step C, x in the mathematical expression is the value at each position on the feature map of the neural network, and a is a learnable parameter. The advantage of this activation function is that it can prevent the gradient from becoming 0 when x is negative, so that the corresponding parameters can continue to be updated.
3. The method for enhancing water degradation images according to claim 1, characterized in that: The attenuation coefficient calculation module in step D includes three groups of 3×3 dilated convolution kernels with dilation rates of 1 / 2 / 5, each group having eight dilated convolution kernels and one group of 3×3 ordinary convolution kernels, with a total of three ordinary convolution kernels.
4. The method for enhancing water degradation images according to claim 1, characterized in that: In the mathematical expression of step E, H, W, and C represent the height, width, and number of channels of the image, respectively, and x, y, and z represent the coordinates of the image in three dimensions. R(I) represents the output of the algorithm in this paper, and R(I) represents the reference target in the training dataset.
5. The method for enhancing water degradation images according to claim 1, characterized in that: In step E, the Adam optimizer is selected for the neural network parameters. The batch size used in deep learning training is 1, the learning rate is 0.001, and the number of training rounds is 1000.