Underwater image enhancement method based on minimum correlation principle and polarization perception enhanced fusion

Abstract

The present application relates to the underwater image enhancement method based on the minimum correlation principle and polarization perception enhanced fusion, compared with the prior art, it solves the defects of insufficient technical fusion, poor scene adaptability, and obvious limitation of enhancement effect.The present application includes the following steps: collecting multiple groups of original underwater polarization images with different polarization angles;generating the brightest polarization image and the darkest polarization image;solving the transmittance and target reflected light;polarization difference and contrast enhancement processing;obtaining the underwater image enhancement result.The present application has stronger parameter estimation stability, can fully exploit the polarization information, and can improve the global and local contrast, and realizes the collaborative optimization of illumination and color.

Description

Technical Field

[0001] This invention relates to the field of image enhancement technology, specifically to an underwater image enhancement method based on the principle of minimum correlation and polarization-sensing enhancement fusion. Background Technology

[0002] With the continuous development of computer vision, image enhancement technology has matured and is widely used in many fields such as marine exploration, underwater robotic operations, and underwater archaeology. It is an important technical means to promote the development of underwater optoelectronic observation equipment. Due to the complex and variable underwater environment, light is prone to attenuation and scattering. Real-world underwater images generally suffer from problems such as color distortion, insufficient clarity, and low contrast, which seriously affect the detection, identification, and tracking of underwater targets, and also hinder the progress of marine environmental research, resource development, and ecological protection.

[0003] Currently, researchers both domestically and internationally have proposed various underwater image enhancement methods, mainly categorized into three types: non-physical model methods, physical model-based methods, and deep learning-based methods. Physical model-based methods rely on prior knowledge to solve parameters for image restoration, closely aligning with underwater imaging principles, but their parameter calculation performance is relatively poor. Deep learning-based methods rely on massive amounts of data to learn image features, achieving better image enhancement results, but they generally suffer from complex model structures and limited practical applications. Non-physical model methods optimize image quality by adjusting pixel distribution, offering ease of use, but they do not incorporate the physical mechanisms of underwater imaging, making them difficult to adapt to complex underwater scenarios.

[0004] Existing traditional enhancement methods mostly have a single processing approach and cannot take into account both underwater optical polarization characteristics and the need for dynamic adjustment of image grayscale. At the same time, few studies combine the minimum correlation principle, polarization difference processing technology and adaptive gamma correction to form a complete processing flow, making it difficult to adapt to the imaging optimization needs of different turbid waters. The overall underwater image enhancement effect has obvious limitations. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies, such as insufficient technology integration, poor scene adaptability, and significant limitations in enhancement effects, by providing an underwater image enhancement method based on the principle of minimum correlation and polarization-sensing enhancement fusion to solve the above problems.

[0006] To achieve the above objectives, the technical solution of the present invention is as follows:

[0007] An underwater image enhancement method based on the principle of minimum correlation and polarization-sensing enhancement fusion includes the following steps:

[0008] Multiple sets of raw underwater polarization images with different polarization angles were acquired;

[0009] Generate the brightest and darkest polarization images: Calculate the Stokes vector based on the original underwater polarization image to obtain the degree of polarization and polarization angle, and generate the brightest and darkest polarization images;

[0010] Solving for transmittance and target reflected light: Construct an adaptive Gaussian kernel and use a Gaussian filter to convolve the original underwater polarization image to separate backscattered light. Based on the backscattered light, simultaneously solve for the transmittance and target reflected light.

[0011] Polarization difference and contrast enhancement processing: Polarization difference processing is performed based on the brightest polarization image and the darkest polarization image, and contrast enhancement is completed step by step by combining transmittance;

[0012] Obtaining underwater image enhancement results: An adaptive gamma correction model is constructed by introducing a brightness perception factor. The gamma correction coefficients are adaptively matched by combining the statistical features of image brightness and the polarization distribution features. Global illumination compensation and color deviation correction are performed on the fused image to output the final enhanced underwater image.

[0013] The generation of the brightest polarization image and the darkest polarization image includes the following steps:

[0014] Based on the original underwater polarization image, the Stokes vector is calculated, and its mathematical expression is as follows:

[0015] ,

[0016] in, It is the Stokes vector we are looking for. , and pixel coordinates The Stokes components at the location, the three of which respectively constitute the total intensity image and the two orthogonal linear polarization component images; , , These are the coordinates when the polarizer is at 0°, 90°, and 45°. The corresponding transmitted light intensity image; x and y are the image pixel plane coordinates;

[0017] By calculating the Stokes vector, we can obtain two key pieces of information about the underwater scene: the degree of polarization and the polarization angle.

[0018] ,

[0019] ,

[0020] in, pixel coordinates Polarization degree DoP, all pixels Combined to form a DoP polarization degree image, pixel coordinates At the polarization angle AoP, = , = , = ;

[0021] Synthesized orthogonal polarization images and This includes the following steps:

[0022] Calculate the light intensity at a rotation angle θ when the light passes completely through the rotating polarizer. It satisfies the following formula:

[0023] ;

[0024] calculate and When polarized light passes through a rotating polarizer, different rotation angles θ correspond to different light intensity outputs, where the rotation angle corresponding to the brightest polarization image is denoted as θ. The rotation angle corresponding to the darkest polarized image is denoted as , and Determined by the following formula: ,

[0025] ,

[0026] Where N is the total number of pixels in the image, the rotation angle of the brightest and darkest polarized images is accurately located by taking the maximum and minimum values ​​of the average light intensity at different angles. For polarizer to turn At an angle, pixels The light intensity of the image taken at that location;

[0027] After determining the key angles, the orthogonal polarization image is directly obtained from the light intensity image at the corresponding angle, i.e.

[0028] ,

[0029] ,

[0030] in, , These are the synthesized orthogonally polarized images, which are also the brightest and darkest polarized images with the highest and lowest average intensities, respectively. These two sets of images serve as the basic operational samples. Polarization angle The corresponding brightest polarization image, Polarization angle The corresponding darkest polarization image.

[0031] The process of determining the transmittance and the target reflected light includes the following steps:

[0032] An adaptive Gaussian kernel is used, which is expressed as:

[0033] ,

[0034] ,

[0035] ,

[0036] in, It is the standard deviation of the Gaussian kernel. and These are the color channel constraint coefficients and the normalized local statistical mean of the entire image; and These are the height and width of the input image. and These are the pixel mean of local window B and the variance of a single pixel in local window B, respectively. It is a very low value to prevent the denominator from being 0; and These are the maximum pixel values ​​for the r, g, and b color channels, respectively. It is a natural index that amplifies the combined effect of the three channel averages to achieve adaptive color adjustment. ;

[0037] Using an adaptive Gaussian kernel, backscattered light and target reflected light are obtained through convolution between a Gaussian filter and the input image. The process is represented as follows:

[0038] ,

[0039] ,

[0040] ,

[0041] in, It is a Gaussian filter. The input color image pixels are RGB. This is an image of backscattered light formed by the scattering of suspended particles in the water. The target reflected light image representing the irradiance of an object after removing the scattering component;

[0042] Since the reflected light and backscattered light from the target satisfy the principle of minimum correlation, their correlation coefficient is calculated using the following formula:

[0043] ,

[0044] ,

[0045] ,

[0046] in, , Reflected light from the target Backscattered light The squared deviation from the image mean at pixel (x,y) represents the light intensity fluctuation of a single pixel. The standard deviation represents the degree of light intensity dispersion in the entire image; The correlation coefficient between the target light and the scattered light under the RGB three channels is the smaller the value, the lower the correlation between the two lights and the better the decomposition effect.

[0047] The iteration uses the calculation formulas from steps 32) and 33) above to iteratively optimize until the constraints are met.

[0048] ,

[0049] Where threshold is the threshold value. , , The correlation coefficients of the target reflected light and the back reflected light in the red, green and blue single channels of the RGB image are shown in order. The maximum value of the correlation coefficient of the three RGB channels is selected as the iterative loss. When the maximum value is less than the threshold, the iteration converges and the decomposition of the two light components is completed.

[0050] After decomposing the target reflected light and backscattered light, the transmittance is obtained by combining the calculated backscattered light. Its expression is:

[0051] ,

[0052] in, It is global backscattered light, estimated using the maximum pixel value of the input image. Transmittance.

[0053] The polarization difference and contrast enhancement processing includes the following steps:

[0054] By performing a difference operation on the brightest polarization image and the darkest polarization image, backscattering noise from the water body is suppressed, resulting in a low-noise polarization difference image and improving the basic contrast of the image.

[0055] By utilizing the difference in polarization properties between the reflected light and the backscattered light from the target: the backscattered light from the water body has a nearly uniform intensity under orthogonal polarization, and is suppressed and eliminated after subtraction; the polarization characteristics of the reflected light from the target are prominent, and the difference is completely preserved. This enables the suppression of scattered light and the extraction of effective target information, providing a foundation for underwater image restoration.

[0056] The intensity distribution of a polarization difference image is calculated using the following expression:

[0057] ,

[0058] in, The intensity distribution of the polarization difference image;

[0059] By leveraging the transmittance adaptive constraint enhancement amplitude, gradient detail enhancement is performed on polarization difference images to enhance target texture and edge features and improve the contrast of local image details.

[0060] The improved gradient is calculated. According to optical imaging theory, the reciprocal of transmittance is used to enhance the gradient field gradient of the detail components, thus obtaining the improved gradient, which is expressed as:

[0061] ,

[0062] Where ∇ represents the gradient operator, and 𝜔 is a non-negative adjustable weight coefficient. To enhance the detail component, For the detail components of the original image, It is the average transmittance;

[0063] Subsequently, the Poisson equation is solved on the improved gradient to further obtain enhanced information, and its expression is:

[0064] ,

[0065] in, To perform the restoration operation for solving the Poisson equation;

[0066] 43) Substitute the transmittance into the atmospheric scattering model to remove image haze interference, and then weight and fuse the detail enhancement component with the dehazed image;

[0067] The overall image contains the total intensity of the underwater scene and provides significant structural information. An atmospheric scattering model is used to remove haze and improve contrast; its expression is:

[0068] ,

[0069] Finally, the dehazed image and the enhanced detail image are merged to generate the enhanced image. Its expression is:

[0070] ,

[0071] in, This is an underwater image after defogging using an atmospheric scattering model. It's transmittance. It is an enhanced image resulting from the fusion of a dehazed image and an image with enhanced details.

[0072] Obtaining the underwater image enhancement result includes the following steps:

[0073] Convert the input DoP polarization degree image from RGB color space to HSV color space using the following expression:

[0074] ,

[0075] HSV stands for Hue, Saturation, and Lightness color space. It is an HSV domain polarization image obtained by spatial mapping of an RGB format DoP polarization image; It is a function for converting RGB to HSV color space. This is an RGB three-channel polarized image generated based on the degree of polarization (DoP), with the three channels corresponding to the red, green, and blue pixel components, respectively.

[0076] The average value of DoP is calculated. The average value of the DoP polarization image is used to characterize the turbidity of the underwater scene. The higher the turbidity of the underwater scene, the lower the average value of DoP. The formula for calculating the average value is as follows:

[0077] ,

[0078] in It is the global mean of all pixels in the entire DoP polarization degree image. It is the DoP polarization degree image pixel value at coordinates (x,y). and These are the height and width of the input image;

[0079] The standard deviation of an image reflects the contrast information of the input image. Using the standard deviation as the basis for designing adaptive gamma coefficients, its calculation formula is as follows:

[0080] ,

[0081] Since higher image turbidity results in lower overall image brightness after dehazing, differentiated gamma correction strategies are designed for different turbidity scenarios:

[0082] A small gamma value is used for brightness compensation in high-turbidity images, and a large gamma value is used for brightness adjustment in low-turbidity images to achieve balanced enhancement of images with different haze levels. The formula for calculating the gamma correction coefficient is designed as follows:

[0083] ,

[0084] in, It is the standard deviation of image pixels. yes The average pixel value of the entire image. Gamma correction factor;

[0085] Based on the calculated gamma correction coefficients, illumination correction is performed on the fused image, and its expression is as follows:

[0086] ,

[0087] in, This is the new luminance channel image after gamma correction.

[0088] A computer-readable storage medium storing a computer program that, when executed by a processor, enables an underwater image enhancement method based on the principle of minimum correlation and polarization-sensing enhancement fusion.

[0089] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, which, when executed by the processor, enables an underwater image enhancement method based on the principle of minimum correlation and polarization-sensing enhancement fusion.

[0090] Beneficial effects

[0091] The underwater image enhancement method of the present invention, based on the principle of minimum correlation and polarization-sensing enhancement fusion, has stronger parameter estimation stability compared with the prior art. It can fully explore polarization information, take into account both global and local contrast improvement, and achieve synergistic optimization of illumination and color.

[0092] The present invention has the following beneficial effects:

[0093] 1. Enhanced contrast by considering both local and global aspects, with stronger adaptive capabilities. This invention relies on polarization difference extraction to extract detail components, which are then weighted and fused with dehazed images from atmospheric models to achieve joint contrast restoration of both local details and global structure, resulting in a natural and stable enhancement effect.

[0094] 2. The separation of target reflected light and backscattered light is more thorough and robust. This invention proposes an iterative decomposition method based on an adaptive Gaussian kernel, which can dynamically adjust the kernel scale according to image features to achieve complete separation of target reflected light and backscattered light. This effectively solves the defect of insufficient separation caused by single decomposition, and significantly improves the stability and accuracy of parameter estimation.

[0095] 3. Fully exploiting polarization information, resulting in outstanding detail restoration. Addressing the issues of insufficient detail and unclear structure in traditional enhancement methods, this invention proposes a multi-dimensional enhancement fusion strategy. This strategy fuses the detailed information of the polarization difference image with the structural information of the total intensity image, effectively overcoming the limitations of a single enhancement method. The enhanced image possesses both clear structure and ample detail.

[0096] 4. Excellent synergistic optimization of illumination and color. This invention introduces a brightness perception factor to construct an adaptive gamma correction model within the HSV variable color separation framework. Utilizing the correlation between polarization degree of polarization (DoP) and water turbidity, the gamma correction coefficient is adaptively and dynamically matched based on the statistical characteristics of the polarization image. The gamma coefficient can be dynamically adjusted according to the mean and standard deviation of the image polarization degree, achieving differentiated illumination compensation for different turbidity scenes. This effectively solves the problem of imbalance between bright and dark areas in traditional enhancement methods, resulting in a corrected image with uniform brightness and natural contrast.

[0097] 5. This invention can provide useful assistance for image segmentation, key point detection and other foggy scenes, and provide a high-quality image foundation for the subsequent processing and analysis of underwater images, which has high practical application value and promotion prospects. Attached Figure Description

[0098] Figure 1 This is a schematic diagram of the process of the present invention;

[0099] Figure 2 This is an overall framework diagram of the underwater image enhancement method of the present invention;

[0100] Figure 3 This is a comparison chart of the present invention with existing methods on the UCCS unpaired dataset;

[0101] Figure 4 This is a comparison chart of the present invention with existing methods on the UIEB pairing dataset;

[0102] Figure 5 This is a comparison chart of the present invention with existing methods on the EUVP unpaired dataset;

[0103] Figure 6 This is a comparison diagram of the present invention with other methods in the restoration of local details;

[0104] Figure 7This diagram compares the application of this invention with other methods in downstream applications of image segmentation and key point detection. Detailed Implementation

[0105] To provide a better understanding of the structural features and effects achieved by the present invention, a detailed description is provided below, accompanied by preferred embodiments and accompanying drawings:

[0106] This invention provides an underwater image enhancement method based on the principle of minimum correlation and polarization-sensing enhancement fusion. The overall process is as follows: Figure 1 As shown, the overall framework is referenced. Figure 2 The structure shown.

[0107] The first step is to acquire multiple sets of raw underwater polarization images with different polarization angles.

[0108] The acquisition of multiple sets of raw underwater images with different polarization angles includes: acquiring three raw images with polarization angles of 0°, 45°, and 90° respectively, forming an image set containing complete polarization information. This invention does not perform any parameter fine-tuning when restoring images taken in different scenarios.

[0109] The second step is to generate the brightest polarization image and the darkest polarization image.

[0110] Based on the obtained raw underwater images, the Stokes vector is calculated. This vector is a fundamental physical quantity describing the polarization state of light, accurately characterizing the polarization distribution of light within the underwater scene. It is the core basis for all subsequent polarization feature calculations, and its mathematical expression is:

[0111] ,

[0112] in, It is the Stokes vector we are looking for. , and pixel coordinates The Stokes components at the location, the three of which respectively constitute the total intensity image and the two orthogonal linear polarization component images; , , These are the coordinates when the polarizer is at 0°, 90°, and 45°. The corresponding transmitted light intensity image; x and y are the image pixel plane coordinates;

[0113] By calculating the Stokes vector, two key pieces of information about the underwater scene can be obtained: the degree of polarization (DoP) and the angle of polarization (AoP). The degree of polarization (DoP) directly reflects the intensity of light polarization, while the angle of polarization (AoP) determines the vibration direction of polarized light. Both are crucial features for distinguishing targets from the water background, and their expressions are as follows:

[0114] ,

[0115] ,

[0116] in, pixel coordinates Polarization degree DoP, all pixels Combined to form a DoP polarization degree image, pixel coordinates At the polarization angle AoP, = , = , = .

[0117] Synthesizing orthogonal polarization images using existing methods and This includes the following steps:

[0118] Calculate the light intensity at a rotation angle θ when the light passes completely through the rotating polarizer. It satisfies the following formula:

[0119] ,

[0120] calculate and When polarized light passes through a rotating polarizer, different rotation angles θ will correspond to different light intensity outputs. The rotation angle corresponding to the brightest polarization image is denoted as θ. The rotation angle corresponding to the darkest polarized image is denoted as And the polarization rotation angle and Determined by the following formula: ,

[0121] ,

[0122] Where N is the total number of pixels in the image, by taking the maximum and minimum values ​​of the average light intensity at different angles, the rotation angle corresponding to the brightest and darkest polarized images can be accurately located. For polarizer to turn At an angle, pixels The light intensity of the image taken at that location;

[0123] After determining the key angles, the orthogonal polarization image can be directly obtained from the light intensity image at the corresponding angle, i.e.

[0124] ,

[0125] ,

[0126] in, , These are the synthesized orthogonally polarized images, which are also the brightest and darkest polarized images with the highest and lowest average intensities, respectively. These two sets of images serve as the basic operational samples. Polarization angle The corresponding brightest polarization image, Polarization angle The corresponding darkest polarization image.

[0127] The third step is to solve for the transmittance and the target reflected light.

[0128] After synthesizing the brightest and darkest polarized images, an iterative target reflection-backscattering light decomposition method is used to separate the target reflected light component from the water backscattering light component, in order to reduce the impact of water scattering on image quality. This includes the following steps:

[0129] By studying the effect of Gaussian kernels on image decomposition, and to overcome the problems of incomplete separation of target reflected light and backscattered light and low robustness, an adaptive Gaussian kernel is adopted, which can be expressed as:

[0130] ,

[0131] ,

[0132] ,

[0133] in, It is the standard deviation of the Gaussian kernel. and These are the color channel constraint coefficients and the normalized local statistical mean of the entire image; and These are the height and width of the input image. and These are the pixel mean of local window B and the variance of a single pixel in local window B, respectively. It is a very low value to prevent the denominator from being 0; and These are the maximum pixel values ​​for the r, g, and b color channels, respectively. It is a natural index that amplifies the combined effect of the three channel averages to achieve adaptive color adjustment. ;

[0134] Using an adaptive Gaussian kernel, backscattered light and target reflected light are obtained through convolution between a Gaussian filter and the input image. This process can be represented as follows:

[0135] ,

[0136] ,

[0137] ,

[0138] in, It is a Gaussian filter. The input color image pixels are RGB. This is an image of backscattered light formed by the scattering of suspended particles in the water. The target reflected light image representing the irradiance of an object after removing the scattering component;

[0139] Since the reflected light and backscattered light from the target satisfy the principle of minimum correlation, their correlation coefficient can be calculated using the following formula:

[0140] ,

[0141] ,

[0142] ,

[0143] in, , Reflected light from the target Backscattered light The squared deviation from the image mean at pixel (x,y) represents the light intensity fluctuation of a single pixel. The standard deviation represents the degree of light intensity dispersion in the entire image; The correlation coefficient between the target light and the scattered light under the RGB three channels is the smaller the value, the lower the correlation between the two lights and the better the decomposition effect.

[0144] The iteration uses the calculation formula from the previous steps to iterate and optimize until the constraints are met.

[0145] ,

[0146] Where threshold is the threshold value. , , The correlation coefficients of the target reflected light and the back reflected light in the red, green and blue single channels of the RGB image are shown in order. The maximum value of the correlation coefficient of the three RGB channels is selected as the iterative loss. When the maximum value is less than the threshold, the iteration converges and the decomposition of the two light components is completed.

[0147] ,

[0148] in, It is global backscattered light, estimated using the maximum pixel value of the input image. Transmittance.

[0149] The fourth step involves performing polarization difference processing on the brightest and darkest polarized images. Based on this polarization difference processing, contrast enhancement is achieved step by step by combining transmittance.

[0150] To overcome the problems of single dimension, insufficient detail, and unclear structure in existing multi-dimensional enhancement and fusion methods for underwater images, this invention employs a multi-dimensional enhancement and fusion strategy to improve the overall visibility of the image. This strategy includes the following steps:

[0151] The difference between the brightest polarization image and the darkest polarization image is calculated to generate polarization difference imaging. The difference in polarization characteristics between backscattered light and target reflected light is used to suppress backscattered light, thereby filtering out some backscattered light.

[0152] This invention can successfully generate light and reflection, while effectively correcting color cast and enhancing contrast and detail.

[0153] The polarization difference image is filtered using the same Gaussian filter as described above to obtain a detailed information image. ,include:

[0154] The intensity distribution of the polarization difference image is calculated using the following expression:

[0155] ,

[0156] in, The intensity distribution of the polarization difference image;

[0157] By leveraging the transmittance adaptive constraint enhancement amplitude, gradient detail enhancement is performed on polarization difference images to enhance target texture and edge features, thereby improving the contrast of local image details.

[0158] The improved gradient is calculated. According to optical imaging theory, the reciprocal of transmittance can be used to enhance the gradient field gradient of the detail components to obtain the improved gradient, which is expressed as follows:

[0159] ,

[0160] Where ∇ represents the gradient operator, It is a non-negative adjustable weighting coefficient. To enhance the amount of detail, For the detail components of the original image, It is the average transmittance;

[0161] Subsequently, the Poisson equation is solved on the improved gradient to further obtain enhanced information, and its expression is:

[0162] ,

[0163] in, Solving the Poisson equation and restoring its original state:

[0164] The detailed information was effectively amplified through the aforementioned enhancements in detail and local contrast.

[0165] By introducing an underwater atmospheric scattering imaging model and combining it with the aforementioned transmittance, the original underwater polarization image is globally dehazed and restored, and the contrast is improved. This eliminates the overall fogging and turbidity effect, resulting in a preliminarily cleared dehazed image.

[0166] The overall image contains the total intensity of the underwater scene and provides significant structural information. However, due to the attenuation effect of light propagation, the overall image exhibits dense fog and low contrast. Therefore, this invention further employs an atmospheric scattering model to remove fog and improve contrast, the expression of which is:

[0167] ,

[0168] Finally, the dehazed image and the enhanced detail image are merged to generate the enhanced image. Its expression is:

[0169] ,

[0170] By integrating the details of the polarization difference image and the structure of the total intensity image, the fused image has a clear structure and ample detail.

[0171] in, This is an underwater image after defogging using an atmospheric scattering model. It's transmittance. It is an enhanced image resulting from the fusion of a dehazed image and an image with enhanced details.

[0172] The fifth step involves introducing a brightness perception factor to construct an adaptive gamma correction model, adaptively matching the gamma correction coefficients, performing global illumination compensation and color deviation correction on the fused image, and outputting the final enhanced underwater image.

[0173] The aforementioned fused image displays good contrast, excellent detail, and a natural appearance, but it suffers from insufficient lighting, resulting in some areas being darker. Traditional fixed-parameter gamma correction cannot match the characteristics of human visual perception, easily leading to localized brightness imbalances. This invention incorporates a brightness perception factor, compensating for attenuated lighting according to the laws of human light perception, and simultaneously correcting color shifts caused by the underwater environment, ultimately outputting an enhanced image with clear layers and true colors.

[0174] This invention further employs an adaptive gamma correction method based on the relationship between turbidity and degree of polarization (DoP) to improve image illumination unevenness. According to underwater polarization optical characteristics, underwater scene turbidity and degree of polarization (DoP) are negatively correlated: as underwater scene turbidity increases, the DoP value gradually decreases. Therefore, the turbidity of the underwater scene can be approximately evaluated using the DoP value, and an adaptive gamma correction strategy can be designed, specifically including the following steps:

[0175] First, convert the RGB color space of the input DoP image to the HSV color space. The expression is as follows:

[0176] ,

[0177] HSV stands for Hue, Saturation, and Lightness color space. It is an HSV domain polarization image obtained by spatial mapping of an RGB format DoP polarization image; It is a function for converting RGB to HSV color space. This is an RGB three-channel polarized image generated based on the degree of polarization (DoP), with the three channels corresponding to the red, green, and blue pixel components, respectively.

[0178] Then, the average value of DoP is calculated. The average value of the DoP image can be used to characterize the turbidity of the underwater scene. The higher the turbidity of the underwater scene, the lower the average value of DoP. The formula for calculating the average value is as follows:

[0179] ,

[0180] in It is the global mean of all pixels in the entire DoP polarization degree image. It is the DoP polarization degree image pixel value at coordinates (x,y). and These are the height and width of the input image;

[0181] The standard deviation of an image reflects the contrast information of the input image. Therefore, this invention uses the standard deviation as the basis for designing the adaptive gamma coefficients, and its calculation formula is as follows:

[0182] ,

[0183] Since higher image turbidity results in lower overall image brightness after dehazing, this invention designs a differentiated gamma correction strategy for different turbidity scenes: using a small gamma value for brightness compensation in high-turbidity images and a large gamma value for brightness adjustment in low-turbidity images, to achieve balanced enhancement of images with different haze levels. Based on the above principle, the gamma correction coefficient calculation formula designed in this invention is as follows:

[0184] ,

[0185] in, It is the standard deviation of image pixels. yes The average pixel value of the entire image. Gamma correction factor;

[0186] Finally, based on the calculated gamma correction coefficients, illumination correction is performed on the fused image, expressed as follows:

[0187] ,

[0188] in, The new luminance channel image after gamma correction;

[0189] The enhanced images have a sharp appearance, good contrast, and ample detail. More importantly, all enhanced images have almost identical brightness, demonstrating the effectiveness and robustness of the adaptive gamma correction method in this invention.

[0190] To comprehensively validate the performance of our proposed method, qualitative and quantitative comparisons were conducted with several state-of-the-art methods on three standard underwater image augmentation datasets (UCCS, UIEB, and EUVP). UCCS and UIEB served as standard test datasets, used to verify the basic capabilities of color correction and comprehensive augmentation of real-world scenes, respectively. EUVP, a large-scale, multi-sample dataset, was used to deeply verify the algorithm's performance in adapting to complex scenes and improving visibility over a wide area. Evaluation metrics used included PCQI, S, and [missing information - likely a specific metric]. Blur and DFAD. PCQI predicts contrast enhancement levels based on human perception mechanisms. 𝑒 and The values ​​represent the enhancement quality in terms of contrast and edges. Blur is typically used to evaluate the blur value of underwater images, ranging from 0 to 1. DFAD evaluates the fog-perceived quality of a given image. To ensure the fairness of the comparative experiments, the method of this invention does not perform any parameter fine-tuning when restoring images taken in different scenes.

[0191] like Figure 3As shown, this graph compares the results of our invention with other methods on the UCCS paired dataset. To demonstrate the color restoration capabilities of our invention, we first qualitatively compare the enhancement results with other methods. From... Figure 3 As can be seen, using ground truth as a reference, some methods used for comparison offer limited quality improvement and fail to effectively correct color casts or remove haze; while others, although significantly improving quality, result in over-enhancement or incorrect color correction. For turbid blue-green underwater images, ShallowNet only corrects the blue-green hue but is ineffective in reducing haze; HAAM, while removing haze in the foreground area, also introduces additional yellow tones; DAUT, LEPFNet, and UShape produce similar results, with insufficient contrast and blurred details; LANet and LiteNet still retain a significant amount of residual haze and have low contrast; HFM can eliminate haze and improve contrast through the effect on some dark areas, but this darkening effect hides some important details. For underwater images with a bluish haze, DAUT and ShallowNet can correct color casts but lack detail; HAAM increases overall brightness but results in distorted color performance; LANet, UShape, and DeepWav still retain severe blue tones. Compared to these methods, the color balance enhancement of the image presented in this invention is natural, significantly improving the natural appearance while enhancing detail and brightness. It exhibits excellent color correction and detail recovery performance, with the enhancement effect almost identical to the ground truth. As can be seen from Table 1, the present invention achieves excellent results on the UCCS verification test group in terms of e, The UCIQE and NIQE metrics are all optimal, edge details are fully restored, and color performance and image naturalness are superior to the comparison algorithms.

[0192] like Figure 4As shown, this figure illustrates a comparison of the present invention with other methods on the UIEB paired dataset. This embodiment provides a broad comparison of the overall enhancement quality on UIEB, with the visual results of different methods shown in the figure. The three original images exhibit slight blue haze, green haze, and severe blue haze, respectively. DAUT and UShape can correct blue tones, but some residual haze remains, and green tones are not significantly corrected; HAAM is effective at removing slight blue and green haze, but ineffective at improving the quality of images with severe blue haze; LANet shows good contrast and rich colors, enhancing slightly bluish-green hazy images, but cannot completely remove severe blue haze; LEPFNet also shows a small amount of residual haze, offering only limited improvement in contrast enhancement and detail recovery; LiteNet has dark areas and loss of detail; PUIENet performs well in correcting color casts, but cannot significantly enhance contrast; ShallowNet slightly corrects color casts, such as the appearance of red tones in ground areas, and is less effective at reducing haze; DeepWav can remove haze to some extent and shows an improvement in contrast, but the result also introduces an undesirable blue tone; HFM significantly eliminates haze throughout the image, however, it creates some dark areas in the enhanced result, which is not conducive to highlighting details. Compared to these methods, this invention achieves enhanced images with realistic color appearance, more complete structure, and prominent details. The colors are true and natural, the contrast is moderate, and the details and textures are clearly highlighted. Its overall enhancement capability is significantly superior to other comparative methods. As shown in Tables 1 and 2, this invention achieves superior enhancement in terms of e, It achieved the best results across all metrics, including UCIQE, NIQE, Blur, and DFAD, balancing detail, color, naturalness, clarity, and dehazing / color correction capabilities.

[0193] like Figure 5As shown, this figure compares the present invention with other methods on the EUVP unpaired dataset. To further evaluate the enhancement performance in diverse underwater environments, comparative experiments were conducted on EUVP. The figure shows that DAUT corrects green and blue-green tones but enhances blue tones and provides a slight enhancement to contrast and detail; HAAM improves the overall brightness of the image without correcting color deviations; LANet and LiteNet perform similarly in enhancing the contrast of hazy blue images but are still insufficient in eliminating unnatural blue tones; LEPFNet and PUIENet are better at improving colors, but their results still show insufficient detail, such as in fish-scale patterns; ShallowNet, UShape, and DeepWav are ineffective in enhancing the visibility and contrast of colorful appearances and have limited enhancement on the other two hazy images; HFM effectively removes haze and significantly improves contrast, but the colors and brightness are unnatural, for example, the background areas introduce residual blue tones. In summary, the enhancement results of the present invention have good appearance, reasonable brightness, and excellent detail recovery capabilities. Combining the data in Tables 1 and 2, it can be seen that the present invention achieved optimal values ​​for all evaluation metrics on the EUVP test set. Specifically, in Table 1, e, ... The metrics demonstrate the advantages of this invention in edge detail restoration, while the UCIQE and NIQE metrics verify the significant improvement in color performance and image naturalness. Table 2 shows that the Blur and DFAD metrics are the lowest, indicating that this invention can effectively reduce image blur, completely remove fog, and correct color deviations. Overall, this invention demonstrates excellent visibility enhancement capabilities and robustness in diverse underwater environments.

[0194] Table 1. Comparison of this invention with other methods on the UCCS, UIEB, and EUVP datasets. UCIQE and NIQE indicator test results

[0195]

[0196] Table 2. Test results of this invention and other methods on the Blur and DFAD metrics of the UCCS, UIEB, and EUVP datasets.

[0197]

[0198] As shown in Figure 6, this invention demonstrates superior performance in contrast and color restoration, significantly enhancing the overall contrast and brightness of the image and clearly restoring its reasonable structure and local details. Comparative experimental results show that the shark image enhanced by ShallowNet is generally blurry, generating mixed results during segmentation; DAUT only corrects the blue-green tones, without significantly improving overall contrast, and the segmentation result is even worse than the original image; HFM enhances local dark details, but leads to unclear segmentation results; HAAM, LEPFNet, and UShape enhance turtle images with similar multi-colored tones, but produce segmentation results with incomplete structures; DeepWav introduces unnecessary blue tones into the enhanced image, resulting in fewer segmented fish. This invention effectively solves the problems of blurred details, insufficient contrast, color deviation, and incomplete structure existing in the above comparative methods, achieving superior enhancement effects in contrast and color richness, providing strong support for segmentation accuracy and the number of segments, and demonstrating outstanding performance in segmentation tasks for targets such as fish and turtles.

[0199] As shown in Figure 7, the enhanced image of this invention provides clearer feature support for keypoint detection. Comparative test results based on the SIFT algorithm show that ShallowNet's enhancement result is blurry, with only a slight increase in the number of keypoints; LANet and LiteNet enhanced images have similar background regions and rich colors, but the number of detected keypoints is almost the same; HAAM and PUIENet have similar keypoint detection locations, with limited improvement in the number of detected keypoints; although HFM significantly improves contrast and color performance, the dark areas generated during the enhancement process limit the number of detected keypoints. Compared with the above methods, the image generated by this invention has a natural color appearance, sufficient detail texture, and reasonable brightness distribution, providing high-quality feature input for keypoint detection. The number of detected keypoints is increased by up to 793.33% compared to the original image, significantly outperforming all comparative methods, verifying the significant performance improvement effect of this invention on downstream image tasks in practical application scenarios.

[0200] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. An underwater image enhancement method based on the principle of minimum correlation and polarization-sensing enhancement fusion, characterized in that, Includes the following steps: 11) Acquire multiple sets of raw underwater polarization images with different polarization angles; 12) Generate the brightest and darkest polarization images: Calculate the Stokes vector based on the original underwater polarization image to obtain the degree of polarization and polarization angle, and generate the brightest and darkest polarization images. 13) Solving for transmittance and target reflected light: Construct an adaptive Gaussian kernel, use a Gaussian filter to convolve the original underwater polarization image to separate backscattered light, and simultaneously solve for transmittance and target reflected light based on the backscattered light. 14) Polarization difference and contrast enhancement processing: Polarization difference processing is performed based on the brightest polarization image and the darkest polarization image, and contrast enhancement is completed step by step by combining transmittance. 15) Obtaining underwater image enhancement results: An adaptive gamma correction model is constructed by introducing a brightness perception factor. The gamma correction coefficient is adaptively matched by combining the statistical features of image brightness and the polarization distribution features. Global illumination compensation and color deviation correction are performed on the fused image to output the final enhanced underwater image.

2. The underwater image enhancement method based on the minimum correlation principle and polarization-sensing enhancement fusion according to claim 1, characterized in that, The generation of the brightest polarization image and the darkest polarization image includes the following steps: 21) Calculate the Stokes vector based on the original underwater polarization image. Its mathematical expression is: ， in, It is the Stokes vector we are looking for. , and pixel coordinates The Stokes components at the location, the three of which respectively constitute the total intensity image and the two orthogonal linear polarization component images; , , These are the coordinates when the polarizer is at 0°, 90°, and 45°. The corresponding transmitted light intensity image; x and y are the image pixel plane coordinates; By calculating the Stokes vector, we can obtain two key pieces of information about the underwater scene: the degree of polarization and the polarization angle. ， ， in, pixel coordinates Polarization degree DoP, all pixels Combined to form a DoP polarization degree image, pixel coordinates At the polarization angle AoP, = , = , = ; 22) Synthesize orthogonal polarization images and This includes the following steps: Calculate the light intensity at a rotation angle θ when the light passes completely through the rotating polarizer. It satisfies the following formula: ； 23) Calculate and When polarized light passes through a rotating polarizer, different rotation angles θ correspond to different light intensity outputs, where the rotation angle corresponding to the brightest polarization image is denoted as θ. The rotation angle corresponding to the darkest polarized image is denoted as , and Determined by the following formula: ， ， Where N is the total number of pixels in the image, the rotation angle of the brightest and darkest polarized images is accurately located by taking the maximum and minimum values ​​of the average light intensity at different angles. For polarizer to turn At an angle, pixels The light intensity of the image taken at that location; After determining the key angles, the orthogonal polarization image is directly obtained from the light intensity image at the corresponding angle, i.e. ， ， in, , These are the synthesized orthogonally polarized images, which are also the brightest and darkest polarized images with the highest and lowest average intensities, respectively. These two sets of images serve as the basic operational samples. Polarization angle The corresponding brightest polarization image, Polarization angle The corresponding darkest polarization image.

3. The underwater image enhancement method based on the minimum correlation principle and polarization-sensing enhancement fusion according to claim 2, characterized in that, The process of determining the transmittance and the target reflected light includes the following steps: 31) An adaptive Gaussian kernel is used, which is expressed as: ， ， ， in, It is the standard deviation of the Gaussian kernel. and These are the color channel constraint coefficients and the normalized local statistical mean of the entire image; and These are the height and width of the input image. and These are the pixel mean of local window B and the variance of a single pixel in local window B, respectively. It is a very low value to prevent the denominator from being 0; and These are the maximum pixel values ​​for the r, g, and b color channels, respectively. It is a natural index that amplifies the combined effect of the three channel averages to achieve adaptive color adjustment. ; 32) Using an adaptive Gaussian kernel, backscattered light and target reflected light are obtained through convolution between a Gaussian filter and the input image. The process is represented as follows: ， ， ， in, It is a Gaussian filter. The input color image pixels are RGB. This is an image of backscattered light formed by the scattering of suspended particles in the water. The target reflected light image representing the irradiance of an object after removing the scattering component; 33) Since the reflected light and backscattered light from the target satisfy the principle of minimum correlation, their correlation coefficient is calculated using the following formula: ， ， ， in, , Reflected light from the target Backscattered light The squared deviation from the image mean at pixel (x,y) represents the light intensity fluctuation of a single pixel. The standard deviation represents the degree of light intensity dispersion in the entire image; The correlation coefficient between the target light and the scattered light under the RGB three channels is the smaller the value, the lower the correlation between the two lights and the better the decomposition effect. The iteration uses the calculation formulas from steps 32) and 33) above to iteratively optimize until the constraints are met. ， Where threshold is the threshold value. , , The correlation coefficients of the target reflected light and the back reflected light in the red, green and blue single channels of the RGB image are shown in order. The maximum value of the correlation coefficient of the three RGB channels is selected as the iterative loss. When the maximum value is less than the threshold, the iteration converges and the decomposition of the two light components is completed. After decomposing the target reflected light and backscattered light, the transmittance is obtained by combining the calculated backscattered light. Its expression is: ， in, It is global backscattered light, estimated using the maximum pixel value of the input image. Transmittance.

4. The underwater image enhancement method based on the minimum correlation principle and polarization-sensing enhancement fusion according to claim 3, characterized in that, The polarization difference and contrast enhancement processing includes the following steps: 41) Perform a difference operation on the brightest polarization image and the darkest polarization image to suppress backscattering noise from the water body, obtain a low-noise polarization difference image, and improve the basic contrast of the image. By utilizing the difference in polarization properties between the reflected light and the backscattered light from the target: the backscattered light from the water body has a nearly uniform intensity under orthogonal polarization, and is suppressed and eliminated after subtraction; the polarization characteristics of the reflected light from the target are prominent, and the difference is completely preserved. This enables the suppression of scattered light and the extraction of effective target information, providing a foundation for underwater image restoration. The intensity distribution of a polarization difference image is calculated using the following expression: ， in, The intensity distribution of the polarization difference image; 42) By relying on the transmittance adaptive constraint to enhance the amplitude, gradient detail enhancement is carried out on the polarization difference image to enhance the target texture and edge features and improve the contrast of local details in the image; The improved gradient is calculated. According to optical imaging theory, the reciprocal of transmittance is used to enhance the gradient field gradient of the detail components, thus obtaining the improved gradient, which is expressed as: ， Where ∇ represents the gradient operator, and 𝜔 is a non-negative adjustable weight coefficient. To enhance the detail component, For the detail components of the original image, It is the average transmittance; Subsequently, the Poisson equation is solved on the improved gradient to further obtain enhanced information, and its expression is: ， in, To perform the restoration operation for solving the Poisson equation; 43) Substitute the transmittance into the atmospheric scattering model to remove image haze interference, and then weight and fuse the detail enhancement component with the dehazed image; The overall image contains the total intensity of the underwater scene and provides significant structural information. An atmospheric scattering model is used to remove haze and improve contrast; its expression is: ， Finally, the dehazed image and the enhanced detail image are merged to generate the enhanced image. Its expression is: ， in, This is an underwater image after defogging using an atmospheric scattering model. It's transmittance. It is an enhanced image resulting from the fusion of a dehazed image and an image with enhanced details.

5. The underwater image enhancement method based on the minimum correlation principle and polarization-sensing enhancement fusion according to claim 4, characterized in that, Obtaining the underwater image enhancement result includes the following steps: 51) Convert the RGB color space of the input DoP polarization image to the HSV color space. The expression is: ， HSV stands for Hue, Saturation, and Lightness color space. It is an HSV domain polarization image obtained by spatial mapping of an RGB format DoP polarization image; It is a function for converting RGB to HSV color space. This is an RGB three-channel polarized image generated based on the degree of polarization (DoP), with the three channels corresponding to the red, green, and blue pixel components, respectively. 52) Calculate the average value of DoP. The average value of the DoP polarization image is used to characterize the turbidity of the underwater scene. The higher the turbidity of the underwater scene, the lower the average value of DoP. The formula for calculating the average value is: ， in It is the global mean of all pixels in the entire DoP polarization degree image. It is the DoP polarization degree image pixel value at coordinates (x,y). and These are the height and width of the input image; 53) The standard deviation of an image reflects the contrast information of the input image. The standard deviation is used as the basis for designing adaptive gamma coefficients, and its calculation formula is as follows: ， Since higher image turbidity results in lower overall image brightness after dehazing, differentiated gamma correction strategies are designed for different turbidity scenarios: A small gamma value is used for brightness compensation in high-turbidity images, and a large gamma value is used for brightness adjustment in low-turbidity images to achieve balanced enhancement of images with different haze levels. The formula for calculating the gamma correction coefficient is designed as follows: ， in, It is the standard deviation of image pixels. yes The average pixel value of the entire image. Gamma correction factor; 54) Based on the calculated gamma correction coefficients, illumination correction is performed on the fused image, expressed as follows: ， in, This is the new luminance channel image after gamma correction.

6. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, enables the underwater image enhancement method based on the principle of minimum correlation and polarization sensing enhancement fusion as described in any one of claims 1-5.

7. A computer device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it can implement the underwater image enhancement method based on the principle of minimum correlation and polarization sensing enhancement fusion as described in any one of claims 1-5.

Images