Underwater image enhancement method and system based on color prior and space-frequency collaborative representation

Abstract

This invention discloses an underwater image enhancement method and system based on color prior and spatial-frequency collaborative representation. The method includes: acquiring an underwater RGB image and extracting basic features; constructing a color degradation prior and generating multi-level guiding information based on the relative attenuation of the blue-red and green-red channels of the underwater image, as well as image saturation and brightness; adaptively modulating the encoded features using scale-adaptive guiding information during the encoding stage, and extracting multi-scale structural information in the spatial domain using a multi-sensory-range parallel branch through a spatial-frequency collaborative feature enhancement unit; performing low-frequency, mid-frequency, and high-frequency soft band division and differential enhancement of the amplitude spectrum in the frequency domain based on learnable frequency band boundary parameters, and fusing spatial-frequency features; and outputting a naturally colored and structurally clear underwater enhanced image through neck integration, multi-level decoding, and skip connections. This invention combines underwater color attenuation prior with spatial-frequency collaborative representation to solve the complex degradation problems of underwater images, such as color shift, structural blurring, and detail attenuation.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and digital image processing technology, specifically to an underwater image enhancement method and system based on color prior and spatial-frequency co-representation. Background Technology

[0002] Underwater images serve as a crucial information source for tasks such as underwater robot operations, marine ecological environment monitoring, and underwater target detection and identification. Their imaging quality directly impacts the reliability of subsequent visual perception and intelligent analysis. However, during propagation through water, light is affected by factors such as absorption and scattering by the water medium, as well as selective attenuation at different wavelengths. This leads to common problems in acquired underwater images, including color distortion, insufficient contrast, blurred edges and texture details, and noise interference. In particular, water attenuates red light far more than blue-green light, causing a common blue-green tint in underwater images. Insufficient brightness and decreased saturation further exacerbate image quality degradation. Therefore, effectively restoring the color, brightness, structure, and details of underwater images has become a critical problem that urgently needs to be solved in this field.

[0003] Early underwater image enhancement methods primarily relied on traditional image processing techniques, physical imaging models, and manually designed prior assumptions to improve image quality. These methods are highly dependent on fixed model assumptions and empirical parameters, limiting their adaptability to the complex and ever-changing underwater environment and making it difficult to consistently achieve ideal enhancement results.

[0004] In recent years, deep learning-based underwater image enhancement methods have made significant progress. Some studies have enhanced color restoration capabilities through attention mechanisms, color space transformation, or multi-branch structures, such as using channel attention to strengthen severely degraded color channels and jointly modeling color information in RGB and HSV spaces. Other studies have introduced frequency domain information to assist in structural or detail restoration, such as fusing Fourier frequency features with spatial features and using gradient correction to enhance edges. Although these methods perform well under certain conditions, they still mainly rely on implicit feature learning or simplified color statistics for color correction. They fail to fully explore and utilize the physical laws of underwater color decay, especially the relative decay relationships between different color channels (such as the relative decay of blue-red and green-red channels) and prior information such as image saturation and brightness. They also fail to explicitly encode these priors into multi-level guidance for adaptive modulation of the encoded features, making it difficult for the network to accurately focus on severely distorted and degraded areas during deep feature learning, resulting in insufficient accuracy and robustness of color correction.

[0005] Meanwhile, while existing space-frequency joint methods attempt to combine spatial and frequency domain features to balance structural and detail restoration, their frequency domain processing mostly employs fixed frequency band divisions or overall spectral transformations. They lack mechanisms for soft separation and differentiated enhancement of different frequency components based on the frequency domain response of underwater image degradation characteristics. Underwater degradation significantly affects low, mid, and high-frequency components: low-frequency components dominate overall color and brightness distribution, mid-frequency components are associated with structural contours and variations, and high-frequency components carry texture details while being affected by noise. If frequency bands cannot be adaptively divided and each band component cannot be enhanced in a targeted manner, problems such as unnatural color restoration, blurred structural contours, or amplified high-frequency noise can easily occur. Therefore, single-domain feature modeling is insufficient to fully characterize the complex composite degradation features of underwater images, and existing space-frequency joint methods still do not adequately explore the intrinsic relationship between underwater degradation characteristics and frequency domain response.

[0006] In summary, there is an urgent need for an enhancement method that can organically combine underwater color attenuation priors with spatial-frequency domain co-representation. This method would utilize color degradation priors to guide feature learning during the encoding process and achieve overall coordinated enhancement of color, brightness, structure, and detail through joint modeling of multi-scale structural information in the spatial domain and multi-band components in the frequency domain. Summary of the Invention

[0007] To address the aforementioned problems, this invention provides an underwater image enhancement method and system based on color priors and spatial-frequency collaborative representation. This invention constructs a color degradation prior including relative attenuation of the blue-red channel, relative attenuation of the green-red channel, image saturation, and brightness, generating multi-level color degradation guidance information. This information is then adaptively modulated to enhance the network's ability to perceive and correct color distortion and brightness degradation, resolving issues such as significant red channel attenuation, prominent blue-green shift, and brightness degradation in underwater images. Simultaneously, this invention sets up parallel feature extraction branches with multiple receptive ranges in the spatial domain to capture structural information at different scales. In the frequency domain, it performs soft band division of the amplitude spectrum into low-frequency, mid-frequency, and high-frequency bands based on learnable frequency band boundary parameters, and applies differentiated enhancement to each frequency band component. Through a spatial-frequency collaborative enhancement mechanism, it achieves overall coordinated restoration of color, brightness, structural contours, and texture details in degraded images, solving the problem of coupled composite degradation features in underwater images and the difficulty in fully representing them using single spatial feature modeling.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides an underwater image enhancement method based on color prior and spatial-frequency co-representation, comprising: Acquire the underwater RGB image to be enhanced; Initial feature extraction is performed on the underwater RGB image to obtain basic features; Constructing a color degradation prior includes: based on the relative attenuation relationship between the blue and red channels and the green and red channels in the underwater RGB image, calculating the relative attenuation information of the blue-red channel and the relative attenuation information of the green-red channel respectively, and converting the underwater RGB image to the HSV color space to obtain saturation information and brightness information. The color degradation prior is formed by the relative attenuation information of the blue-red channel, the relative attenuation information of the green-red channel, the saturation information, and the brightness information. Based on the color degradation prior, color degradation guidance information is generated for different coding levels; The basic features are sequentially input into multi-level coding processing, where adjacent coding levels are connected by downsampling. In each level of coding processing, color degradation guidance information adapted to the spatial scale of the current coding level is used to adaptively modulate the coding features, and the coding features are jointly enhanced in the spatial and frequency domains through space-frequency collaborative feature enhancement processing. The joint enhancement in the frequency domain includes: performing Fourier transform on the coding features to obtain amplitude spectrum and phase spectrum; dividing the amplitude spectrum into low-frequency, mid-frequency, and high-frequency soft bands based on learnable frequency band boundary parameters; enhancing the amplitude spectrum of each frequency band separately; and combining it with the phase spectrum to perform inverse Fourier transform to obtain frequency domain enhanced features. The deep features obtained through the multi-level encoding process are integrated, and the integrated features are then subjected to multi-level decoding. Spatial information is recovered through upsampling, and features are supplemented by skip connections between the encoding and decoding processes. Finally, an enhanced underwater image is output.

[0009] In one embodiment of the present invention, the color degradation prior... Represented as:

[0010] in, This refers to the relative attenuation information of the blue and red channels. S represents the relative attenuation information of the green and red channels, S represents saturation information, V represents brightness information, and B, G, and R represent the values ​​of the blue, green, and red channels, respectively. To prevent extremely small positive numbers with a denominator of zero.

[0011] In one embodiment of the present invention, the step of generating color degradation guidance information for different coding levels based on the color degradation prior, and the adaptive modulation, include: The color degradation prior is subjected to convolutional mapping, nonlinear activation and channel compression to obtain the color degradation guidance map; Based on the spatial resolution of features at different coding levels, the color degradation guide map is scale-adapted to obtain a result consistent with the first... Guided features with consistent feature space size at each coding level ; For the guiding features Perform guided feature mapping operation The modulation weights are obtained through the Sigmoid function. The modulation weights are then compared with the input features of the current coding level. Multiplying element by element yields the modulated features.

[0012] In one embodiment of the present invention, the guided feature mapping operation This includes convolution operations and non-linear activation.

[0013] In one embodiment of the present invention, the space-frequency cooperative feature enhancement processing includes: The input features are normalized and processed separately through spatial attention and channel attention branches. The results of the two branches are then fused to obtain the basic enhanced features. The basic enhancement features are subjected to spatial domain enhancement processing, which includes mapping the basic enhancement features to a low-resolution feature space, extracting multi-scale information through multiple parallel feature extraction branches with different receptive ranges, fusing the multi-scale information and restoring the spatial scale to obtain spatial domain enhancement features.

[0014] In one embodiment of the present invention, the soft band division process is represented as follows:

[0015] in, Represents the amplitude spectrum of the input features. , and These represent the low-frequency amplitude spectrum, mid-frequency amplitude spectrum, and high-frequency amplitude spectrum, respectively. This represents a soft band partitioning operation based on learnable frequency band boundary parameters. and This represents the learnable frequency band boundary parameters.

[0016] In one embodiment of the present invention, the process of enhancing the amplitude spectra of each frequency band obtained by division and then combining them with the phase spectrum to perform an inverse Fourier transform is expressed as follows:

[0017] in, Indicates the inverse Fourier transform. Represents the original phase spectrum. , and These represent low-frequency enhancement processing, mid-frequency enhancement processing, and high-frequency enhancement processing, respectively. This represents the enhanced features recovered after frequency domain processing.

[0018] In one embodiment of the present invention, the plurality of parallel feature extraction branches with different receptive ranges are implemented using depthwise separable convolutions with different dilation rates.

[0019] In one embodiment of the present invention, the multi-level encoding process includes a first-level encoding process, a second-level encoding process, and a third-level encoding process connected in sequence, with each level of encoding process connected to the next level through a downsampling operation.

[0020] Secondly, the present invention provides an underwater image enhancement system based on color prior and spatial-frequency co-representation, used to implement the underwater image enhancement method based on color prior and spatial-frequency co-representation, wherein the underwater image enhancement system based on color prior and spatial-frequency co-representation includes: The initial feature extraction unit is used to extract initial features from the input underwater RGB image to obtain basic features; An underwater color degradation perception unit is used to construct a color degradation prior based on the relative attenuation relationship between different color channels of the underwater RGB image and the saturation and brightness information of the image, and to generate color degradation guidance information for different coding levels; wherein, the construction of the color degradation prior includes calculating the relative attenuation information of the blue-red channel and the relative attenuation information of the green-red channel, and converting the image to the HSV color space to obtain saturation and brightness information; The multi-level coding processing unit includes multiple cascaded coding processing units and downsampling units disposed between adjacent coding processing units. Each coding processing unit receives the output features of the previous level, adaptively modulates the input features using color degradation guidance information adapted to the spatial scale of the current coding level, and performs spatial and frequency domain collaborative enhancement processing on the input features through a built-in space-frequency collaborative feature enhancement unit to output coded features. The space-frequency collaborative feature enhancement unit includes a subunit for dividing the amplitude spectrum into low-frequency, mid-frequency, and high-frequency soft bands based on learnable frequency band boundary parameters in the frequency domain. The neck processing unit is used to integrate the deep features output by the multi-level coding processing unit; A multi-level decoding processing unit includes multiple decoded processing units connected in series and an upsampling unit disposed between adjacent decoded processing units; the upsampling unit is used to upsample the output features of the previous level, and the decoded processing unit is used to fuse the upsampled features with the features introduced by the corresponding encoding processing unit through a skip connection, and recover the spatial information. An output mapping unit is used to generate an enhanced underwater image based on the features output by the multi-level decoding processing unit.

[0021] Compared with the prior art, the present invention has at least the following beneficial effects: 1. This invention constructs a color degradation prior that includes relative attenuation of blue-red channels, relative attenuation of green-red channels, image saturation, and brightness, and generates multi-level color degradation guidance information to adaptively modulate the coding features. This enables the network to effectively focus on areas of severe color distortion and brightness degradation at different coding levels, significantly improving the perception and correction capabilities of underwater image color shift and insufficient brightness, resulting in more accurate and natural color restoration.

[0022] 2. This invention sets up multiple parallel feature extraction branches with different receptive ranges in the spatial domain to capture local, medium and large-scale structural context information, which effectively enhances the network's ability to recover structural contours at different scales and improves the structural blurring problem caused by scattering and attenuation in underwater images.

[0023] 3. This invention performs soft band division of the amplitude spectrum into low-frequency, mid-frequency, and high-frequency bands in the frequency domain using learnable band boundary parameters, and applies differentiated enhancement to each band component: the low-frequency component focuses on restoring overall brightness and color, the mid-frequency component focuses on restoring information related to structural changes, and the high-frequency component focuses on restoring texture details and suppressing noise. This mechanism enables frequency domain enhancement to adaptively and finely process different degradation features, overcoming the shortcomings of insufficient adaptability of fixed band division, and better balancing structure preservation and detail enhancement.

[0024] 4. This invention fuses spatial domain enhancement features with frequency domain enhancement features, achieving synergistic representation of spatial structural information and frequency domain characteristics. This enables the network to jointly model the composite degradation features of underwater images, effectively solving the problem that color cast, structural blur, contrast reduction, and detail attenuation are mutually coupled and difficult to fully represent by single-domain feature modeling. The output enhanced image achieves overall coordinated restoration in color, brightness, and structural details.

[0025] 5. Test results on publicly available underwater image enhancement benchmark datasets show that the method of this invention achieves the best performance in objective indicators such as peak signal-to-noise ratio, structural similarity, and learned perceptual image patch similarity. The enhancement results are closest to the real labeled images, and the subjective visual effect is also significantly better than existing methods, verifying the advanced nature and effectiveness of the technical solution of this invention. Attached Figure Description

[0026] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the drawings without creative effort.

[0027] Figure 1 This is an overall flowchart of the method of the present invention.

[0028] Figure 2 This is a schematic diagram of the color degradation sensing unit of the present invention.

[0029] Figure 3 This is a schematic diagram of the space-frequency cooperative feature enhancement unit of the present invention.

[0030] Figure 4 This is a qualitative comparison of the method of this invention (Ours) with existing methods on the UIEB dataset. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited to the following embodiments. For those skilled in the art, equivalent substitutions, combinations or conventional modifications made without departing from the concept of the present invention should all fall within the scope of protection of the present invention.

[0032] like Figure 1 , Figure 2 and Figure 3 As shown, this invention provides an underwater image enhancement method based on color prior and spatial-frequency co-representation, comprising the following steps: Step 1: Input the underwater RGB image to be enhanced into the underwater image enhancement network to obtain basic features through initial feature extraction; at the same time, input the underwater RGB image to be enhanced into the color degradation perception unit, construct color degradation priors based on the relative attenuation relationship between different color channels and the saturation and brightness information of the image, and generate color degradation guidance information for different coding levels. Step 2: The basic features are sequentially input into the multi-level coding processing unit, and adjacent coding processing units are connected through downsampling operations. During each level of coding, color degradation guidance information adapted to the spatial scale of the current coding level is used to adaptively modulate the coding features. The spatial and frequency domain collaborative processing of the coding features is performed by the space-frequency collaborative feature enhancement unit set in each level of coding processing unit to enhance the network's ability to represent composite degradation features such as color distortion, structural blurring, and texture detail attenuation. Step 3: The deep features obtained after multi-level encoding are input into the neck processing unit for integration, and then sequentially input into the multi-level decoding processing unit. The image spatial information is gradually recovered through upsampling operations. The effective features retained in the encoding stage are supplemented by the skip connections between the encoding processing unit and the corresponding decoding processing unit. Finally, the enhanced result image is obtained through output mapping.

[0033] In one embodiment of the present invention, the color degradation sensing unit in step 1 includes color degradation prior construction, prior feature mapping and multi-scale guided modulation process.

[0034] The color degradation prior construction process includes: obtaining the red channel R, green channel G, and blue channel B from the input underwater RGB image, and calculating the relative attenuation information of the blue and red channels. Relative attenuation information between green and red channels Simultaneously, the input RGB image is converted to the HSV color space to obtain saturation information S and brightness information V. The color degradation prior is represented as follows:

[0035] In the formula, This indicates a priori color degradation. This indicates the relative attenuation information of the blue and red channels. This represents the relative attenuation information of the green and red channels, S represents saturation information, and V represents brightness information. This refers to an extremely small positive number used to prevent the denominator from being zero and to improve the stability of numerical calculations. In one embodiment of the invention, .

[0036] After obtaining the color degradation prior, it is input into the prior feature mapping module for feature integration to generate a color degradation guidance map. The prior feature mapping module uses convolutional mapping, nonlinear activation, and channel compression operations to process the multi-channel color degradation prior, thereby obtaining color degradation guidance features suitable for encoding feature modulation.

[0037] In one embodiment of the present invention, the color degradation guide map is scale-adapted according to the spatial resolution of different coding level features to obtain a result consistent with the first... Guided features with consistent feature space size at each coding level Subsequently, modulation weights are generated based on the guiding features, and these modulation weights are applied to the input features of the current coding level. The modulated output characteristics are obtained. The process is represented as follows:

[0038] In the formula, This indicates a guided feature mapping operation. This represents the Sigmoid function. This indicates element-wise multiplication. Through the above processing, the color degradation state of the input underwater image is incorporated into the adaptive modulation of different levels of encoded features.

[0039] In one embodiment of the present invention, step 2 sets up a first-level encoding processing unit, a second-level encoding processing unit, and a third-level encoding processing unit, with adjacent encoding processing units connected via a downsampling operation. The first-level, second-level, and third-level encoding processing units respectively receive color degradation guidance information adapted to the feature space resolution of the current level, and each level of encoding processing unit sets up a space-frequency collaborative feature enhancement unit.

[0040] The spatial-frequency collaborative feature enhancement unit first normalizes the input features, and then models the importance of different spatial locations and the response weights of different feature channels through spatial attention and channel attention branches, respectively. The processing results of the two attention branches are fused to obtain the basic enhanced features for subsequent spatial and frequency domain processing.

[0041] In the spatial domain processing, the basic enhanced features are first mapped to a low-resolution feature space. Then, multiple parallel feature extraction branches are set up in the low-resolution feature space, each branch employing feature extraction operations with different receptive ranges to extract local, medium-range, and large-scale contextual information, respectively. In one embodiment of the invention, the parallel feature extraction branches are implemented using depthwise separable convolutions with different dilation rates. After fusion mapping and spatial scale recovery, the output features of multiple scale branches yield spatial domain enhanced features.

[0042] In the frequency domain processing, the input features are first mapped to the frequency domain using a two-dimensional fast Fourier transform to obtain the amplitude spectrum and phase spectrum. Then, based on the radial distance of the frequency domain position relative to the spectrum center, and combined with learnable frequency band boundary parameters, the amplitude spectrum is divided into low-frequency, mid-frequency, and high-frequency soft bands. This process is represented as follows:

[0043] In the formula, Represents the amplitude spectrum of the input features. , and These represent the low-frequency amplitude spectrum, mid-frequency amplitude spectrum, and high-frequency amplitude spectrum, respectively. This represents a soft band partitioning operation based on learnable frequency band boundary parameters. and This represents the learnable frequency band boundary parameters.

[0044] In one embodiment of the present invention, the low-frequency amplitude spectrum is used to enhance overall brightness and color information, the mid-frequency amplitude spectrum is used to enhance structural change-related information, and the high-frequency amplitude spectrum is used to enhance texture details and suppress noise interference. Differential enhancement processing is performed on the low-frequency, mid-frequency, and high-frequency amplitude spectra respectively, and the enhanced frequency band information is combined with the original phase information to restore spatial domain features. The process is represented as follows:

[0045] In the formula, Indicates the inverse Fourier transform. Represents the original phase spectrum. , and These represent low-frequency enhancement processing, mid-frequency enhancement processing, and high-frequency enhancement processing, respectively. This represents the enhanced features recovered after frequency domain processing.

[0046] Subsequently, the spatial domain enhancement features and frequency domain enhancement features are fused and mapped to obtain the spatial-frequency co-enhanced features of the current coding level, which are then passed to subsequent network levels.

[0047] In one embodiment of the present invention, step 3 involves integrating the deep features obtained after multi-level encoding processing into the neck processing unit, and then inputting the integrated features into the decoding stage. The decoding stage gradually recovers the image spatial information through multi-level decoding processing and upsampling operations; skip connections are established between the encoding processing unit and the corresponding decoding processing unit to transmit the effective features retained in the encoding stage. Finally, the decoded features are mapped to generate an enhanced result image.

[0048] The effects of this invention can be further illustrated by the following experiments.

[0049] Dataset: This invention uses the Underwater Image Enhancement Benchmark (UIEB) dataset. UIEB is a recognized authoritative benchmark in the field, containing 890 paired images. In this experiment, it was randomly divided into 800 training images and 90 test images.

[0050] Evaluation metrics: Peak signal-to-noise ratio (PSNR) is selected to measure the reconstruction accuracy of the enhanced image and the reference image at the pixel level. Structural similarity (SSIM) is used to evaluate the algorithm's ability to preserve image brightness, contrast and structural information. Learning perceptual patch similarity (LPIPS) is introduced to evaluate the degree of conformity between the enhancement results and human visual perception in a higher-dimensional semantic space. Training Setup: This invention implements the underwater image enhancement network based on the PyTorch deep learning framework and uses a single NVIDIA GeForce RTX 5080 GPU for training and testing. During training, the input underwater images and their corresponding reference images are uniformly adjusted to 256×256 pixels, and the batch size is set to 4. The network parameters are updated using the Adam optimizer, and the initial learning rate is set to 1×10⁻⁶. -4Furthermore, a cosine annealing strategy is employed to dynamically adjust the learning rate, allowing it to gradually change throughout the training process, thereby improving the stability and convergence performance of network training.

[0051] Loss Function: To simultaneously ensure the reconstruction accuracy at the pixel level and the perceptual consistency in the high-level feature space, this embodiment uses a hybrid loss function consisting of pixel-level L1 loss and perceptual loss to optimize the network. The total loss function is expressed as:

[0052] in, and These represent the network's output image and the real image, respectively. express Norm. express Feature extractor: In this invention, the weights of the perceptual loss are set in the experiment. The value is 0.01. Pixel-level L1 loss is used to constrain the pixel differences between the output image and the reference image to maintain the consistency of overall content and color; perceptual loss is used to constrain the differences between the two in the feature space to enhance the recovery of image structure, edges and high-frequency texture details.

[0053] Comparison Methods: Several representative benchmark models were selected for comparison, including the early generative adversarial network-based FUnIE-GAN, the UIEC^2-Net which integrates color correction and enhancement strategies, the PUIE which adopts a probability distribution estimation strategy, and the SFG-Net, CCL-Net, CDF-UIE and Polykernel which have performed well in feature extraction and degradation modeling in recent years.

[0054] Results Analysis: Table 1 shows the quantitative comparison results of our proposed method (Ours) with existing methods on the UIEB dataset. Our proposed method achieves the best PSNR, SSIM, and LPIPS on the UIEB dataset, and its overall performance remains highly competitive. Our advantage stems from our targeted design for underwater composite degradation: First, the color degradation perception unit explicitly enhances the model's ability to perceive and correct color distortion regions, thereby improving the accuracy and stability of color restoration; second, the space-frequency co-enhancement unit, by jointly modeling spatial structure information and frequency degradation features, enables the network to restore overall tone consistency while also considering local structure and texture details, avoiding the problem of only increasing brightness while neglecting detail preservation.

[0055] Table 1. Quantitative comparison results of the method of this invention (Ours) and existing methods on the UIEB dataset.

[0056]

[0057] from Figure 4 It can be seen that the enhancement result of the method of this invention is closest to the real label image. Among the comparison methods, FUnIE-GAN, SFG-Net, and CDF-UIE still exhibit significant color casts in some samples. Specifically, FUnIE-GAN's restored result shows inconsistent tonal variations; SFG-Net's overall tone is cool and lacks contrast; and CDF-UIE shows a yellowish-green tendency in its restoration. In contrast, UIEC^2-Net and Polykernel have relatively natural overall tones, but still suffer from saturation distortion and lack sufficient visual depth. PUIE's output has a certain hazy quality, lacking clarity and transparency. CCL-Net performs relatively stably in color restoration, but brightness and local details are still insufficient.

[0058] Furthermore, in some embodiments, the present invention also provides an underwater image enhancement system based on color prior and spatial-frequency co-representation, used to implement the underwater image enhancement method based on color prior and spatial-frequency co-representation, wherein the underwater image enhancement system based on color prior and spatial-frequency co-representation includes: The initial feature extraction unit is used to extract initial features from the input underwater RGB image to obtain basic features; An underwater color degradation perception unit is used to construct a color degradation prior based on the relative attenuation relationship between different color channels of the underwater RGB image and the saturation and brightness information of the image, and to generate color degradation guidance information for different coding levels; wherein, the construction of the color degradation prior includes calculating the relative attenuation information of the blue-red channel and the relative attenuation information of the green-red channel, and converting the image to the HSV color space to obtain saturation and brightness information; The multi-level coding processing unit includes multiple cascaded coding processing units and downsampling units disposed between adjacent coding processing units. Each coding processing unit receives the output features of the previous level, adaptively modulates the input features using color degradation guidance information adapted to the spatial scale of the current coding level, and performs spatial and frequency domain collaborative enhancement processing on the input features through a built-in space-frequency collaborative feature enhancement unit to output coded features. The space-frequency collaborative feature enhancement unit includes a subunit for dividing the amplitude spectrum into low-frequency, mid-frequency, and high-frequency soft bands based on learnable frequency band boundary parameters in the frequency domain. The neck processing unit is used to integrate the deep features output by the multi-level coding processing unit; A multi-level decoding processing unit includes multiple decoded processing units connected in series and an upsampling unit disposed between adjacent decoded processing units; the upsampling unit is used to upsample the output features of the previous level, and the decoded processing unit is used to fuse the upsampled features with the features introduced by the corresponding encoding processing unit through a skip connection, and recover the spatial information. An output mapping unit is used to generate an enhanced underwater image based on the features output by the multi-level decoding processing unit.

[0059] In summary, the underwater image enhancement method and system based on color prior and spatial-frequency co-representation provided by this invention constructs a color degradation prior including relative attenuation of blue-red channels, relative attenuation of green-red channels, image saturation, and brightness. It then generates multi-level color degradation guidance information to adaptively modulate the coding features at each level, enabling the network to effectively focus on areas of color distortion and brightness degradation. The spatial-frequency co-representation feature enhancement unit, set in the coding stage, extracts multi-scale structural information in the spatial domain through parallel branches with multiple receptive ranges. In the frequency domain, it performs soft band division of the amplitude spectrum based on learnable frequency band boundary parameters and implements differentiated enhancement for each frequency band component, fusing the spatial and frequency domain enhancement features to achieve the coordinated restoration of color, structure, and texture details. After neck integration and multi-level decoding, skip connections supplement the effective features retained in the coding stage, ultimately generating an underwater enhanced image with natural colors, good contrast, and clear details. This invention effectively solves the complex degradation problems of color shift, structural blurring, and detail attenuation in underwater images, significantly outperforming existing methods in both quantitative evaluation and visual quality.

[0060] Although the invention has been described with reference to preferred embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, the technical features mentioned in the various embodiments can be combined in any manner as long as there is no structural conflict. The invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. An underwater image enhancement method based on color priors and space-frequency collaborative representation, characterized in that, include: Acquire the underwater RGB image to be enhanced; Initial feature extraction is performed on the underwater RGB image to obtain basic features; Constructing a color degradation prior includes: based on the relative attenuation relationship between the blue and red channels and the green and red channels in the underwater RGB image, calculating the relative attenuation information of the blue-red channel and the relative attenuation information of the green-red channel respectively, and converting the underwater RGB image to the HSV color space to obtain saturation information and brightness information. The color degradation prior is formed by the relative attenuation information of the blue-red channel, the relative attenuation information of the green-red channel, the saturation information, and the brightness information. Based on the color degradation prior, color degradation guidance information is generated for different coding levels; The basic features are sequentially input into multi-level coding processing, where adjacent coding levels are connected by downsampling. In each level of coding processing, color degradation guidance information adapted to the spatial scale of the current coding level is used to adaptively modulate the coding features, and the coding features are jointly enhanced in the spatial and frequency domains through space-frequency collaborative feature enhancement processing. The joint enhancement in the frequency domain includes: performing Fourier transform on the coding features to obtain amplitude spectrum and phase spectrum; dividing the amplitude spectrum into low-frequency, mid-frequency, and high-frequency soft bands based on learnable frequency band boundary parameters; enhancing the amplitude spectrum of each frequency band separately; and combining it with the phase spectrum to perform inverse Fourier transform to obtain frequency domain enhanced features. The deep features obtained through the multi-level encoding process are integrated, and the integrated features are then subjected to multi-level decoding. Spatial information is recovered through upsampling, and features are supplemented by skip connections between the encoding and decoding processes. Finally, an enhanced underwater image is output. The color degradation prior is represented as: wherein, is the relative attenuation information for the blue-red channel, is the relative attenuation information for the green-red channel, S is the saturation information, V is the luminance information, B, G, R are the blue, green, red channel values, respectively, is a small positive number to prevent division by zero. The step of generating color degradation guidance information for different coding levels based on the color degradation prior, and the adaptive modulation, include: The color degradation prior is subjected to convolutional mapping, nonlinear activation and channel compression to obtain the color degradation guidance map; The color degradation guide map is scale adapted according to spatial resolutions of different coding level features to obtain guide features consistent with spatial sizes of the first coding level features ; to the guiding feature performing a guiding feature mapping operation and get the modulation weight through the Sigmoid function; modulating the input features of the current coding level by the modulation weights element-wise multiplication, resulting in modulated features.

2. The method of underwater image enhancement based on color priors and space-frequency collaborative representation according to claim 1, characterized in that, The guided feature mapping operation includes a convolution operation and a nonlinear activation.

3. The method of underwater image enhancement based on color priors and space-frequency collaborative representation according to claim 1, characterized in that, The space-frequency cooperative feature enhancement process includes: The input features are normalized and processed separately through spatial attention and channel attention branches. The results of the two branches are then fused to obtain the basic enhanced features. The basic enhancement features are subjected to spatial domain enhancement processing, which includes mapping the basic enhancement features to a low-resolution feature space, extracting multi-scale information through multiple parallel feature extraction branches with different receptive ranges, fusing the multi-scale information and restoring the spatial scale to obtain spatial domain enhancement features.

4. The method of underwater image enhancement based on color priors and space-frequency collaborative representation according to claim 1, characterized in that, The process of soft band division is represented as follows: wherein, denotes an input feature, , and denote a low-frequency amplitude spectrum, a mid-frequency amplitude spectrum, and a high-frequency amplitude spectrum, respectively, denotes a soft frequency band division operation based on a learnable frequency band boundary parameter, and denote a learnable frequency band boundary parameter.

5. The method of underwater image enhancement based on color priors and space-frequency collaborative representation according to claim 4, characterized in that, The process of enhancing the amplitude spectrum of each frequency band obtained by division and then combining it with the phase spectrum to perform an inverse Fourier transform is expressed as follows: wherein, denotes an inverse Fourier transform, denotes an original phase spectrum, , and denote a low frequency enhancement processing, a mid frequency enhancement processing and a high frequency enhancement processing, respectively, denotes an enhanced feature recovered after the frequency domain processing.

6. The method of underwater image enhancement based on color priors and space-frequency collaborative representation according to claim 3, characterized in that, The multiple parallel feature extraction branches with different receptive ranges are implemented using depthwise separable convolutions with different dilation rates.

7. The method of underwater image enhancement based on color priors and space-frequency collaborative representation according to claim 1, characterized in that, The multi-level encoding process includes a first-level encoding process, a second-level encoding process, and a third-level encoding process connected in sequence. After each level of encoding process, the next level is connected through a downsampling operation.

8. An underwater image enhancement system based on color prior and spatial-frequency co-representation, characterized in that, A method for implementing claim 1 7. The method of underwater image enhancement based on color priors and spatio- spectral collaborative representation of any one of the preceding claims, the system of underwater image enhancement based on color priors and spatio-spectral collaborative representation comprising: The initial feature extraction unit is used to extract initial features from the input underwater RGB image to obtain basic features; An underwater color degradation perception unit is used to construct a color degradation prior based on the relative attenuation relationship between different color channels of the underwater RGB image and the saturation and brightness information of the image, and to generate color degradation guidance information for different coding levels; wherein, the construction of the color degradation prior includes calculating the relative attenuation information of the blue-red channel and the relative attenuation information of the green-red channel, and converting the image to the HSV color space to obtain saturation and brightness information; The multi-level coding processing unit includes multiple cascaded coding processing units and downsampling units disposed between adjacent coding processing units. Each coding processing unit receives the output features of the previous level, adaptively modulates the input features using color degradation guidance information adapted to the spatial scale of the current coding level, and performs spatial and frequency domain collaborative enhancement processing on the input features through a built-in space-frequency collaborative feature enhancement unit to output coded features. The space-frequency collaborative feature enhancement unit includes a subunit for dividing the amplitude spectrum into low-frequency, mid-frequency, and high-frequency soft bands based on learnable frequency band boundary parameters in the frequency domain. The neck processing unit is used to integrate the deep features output by the multi-level coding processing unit; A multi-level decoding processing unit includes multiple decoded processing units connected in series and an upsampling unit disposed between adjacent decoded processing units; the upsampling unit is used to upsample the output features of the previous level, and the decoded processing unit is used to fuse the upsampled features with the features introduced by the corresponding encoding processing unit through a skip connection, and recover the spatial information. An output mapping unit is used to generate an enhanced underwater image based on the features output by the multi-level decoding processing unit.

Images