Method for non-reference underwater image enhancement based on morphological intelligence and sparse representation

Abstract

The application discloses a method for underwater image enhancement based on morphological intelligence and sparse representation, and relates to the technical field of underwater image enhancement, and comprises the following steps: S1, adaptive nonlinear white balance: a nonlinear white balance correction method based on data driving is adopted to dynamically adjust a global gain and correct color deviation caused by non-uniform illumination; S2, morphological contrast enhancement: a channel-level processing method is adopted to combine top-hat transformation and bottom-hat transformation, enhance global contrast, and highlight hidden targets on the premise of retaining macroscopic structure; and S3, sparse representation detail enhancement: a non-subsample contourlet (NSCT) is adopted to construct a super-complete dictionary, cross-scale consistency features are utilized to extract fine targets and suppress scattering noise. The method for underwater image enhancement based on morphological intelligence and sparse representation can accurately distinguish noise and real details, realize clarity and texture restoration effect, and is suitable for image defogging, key point detection and edge detection.

Description

Technical Field

[0001] This invention relates to the field of underwater image enhancement technology, and in particular to a referenceless underwater image enhancement method based on morphological intelligence and sparse representation. Background Technology

[0002] With the advancement of the maritime power strategy, the demand for high-definition underwater visual information is becoming increasingly urgent to support precise perception and intelligent operation in the deep-sea environment. Acquiring high-quality underwater images plays a crucial role in improving AUV navigation accuracy, target recognition capabilities, and mission performance. However, underwater imaging is inevitably affected by light absorption and scattering, often leading to problems such as color cast (e.g., blue-green hues), reduced contrast, and blurred details. These imaging degradations, caused by water composition, imaging distance, and lighting conditions, severely impair the reliability of visual systems and the accuracy of advanced tasks such as obstacle detection and target recognition. Therefore, many researchers are dedicated to the development of color correction and image enhancement technologies. However, existing methods often produce new artifacts (such as overcorrection or edge detail smoothing), resulting in image blurring. Solving these complex imaging degradation problems remains a core challenge in the current development of underwater vision technology.

[0003] Methods based on non-physical models significantly improve visual quality by employing spatial computations (such as gamma correction and histogram mapping) and frequency domain transformations (such as wavelet analysis and Fourier analysis). However, these methods encounter fundamental bottlenecks when applied to complex underwater environments. The main obstacle lies in the significant spatial non-uniformity and depth-dependent characteristics of underwater optical degradation. Specifically, variations in optical path length induce heterogeneous absorption and scattering effects at different depths, making the degradation process inherently nonlinear and difficult to model using global linear operators.

[0004] Existing spatial domain methods, while primarily based on the local neighborhood assumption and effectively enhancing local contrast, struggle to maintain global illumination consistency, often resulting in color banding or brightness discontinuities. Conversely, frequency domain methods, while possessing global modeling capabilities, often fail to distinguish between real texture and scattering noise due to their strategy of applying uniform gain to high-frequency components—a problem particularly pronounced in weak signal regions, leading to semantic information loss or excessive smoothing of details. Furthermore, single-domain processing struggles to balance multi-dimensional requirements such as color fidelity, structural contrast, and texture sharpening, as these enhancement objectives interfere with each other due to operational coupling. Therefore, an enhancement framework with cross-domain collaboration, hierarchical decoupling, and adaptive adjustment capabilities is urgently needed. This framework must accurately restore local details and address the nonlinear degradation caused by depth variations, without compromising overall consistency. Summary of the Invention

[0005] The purpose of this invention is to provide a referenceless underwater image enhancement method based on morphological intelligence and sparse representation, which solves the problems of color deviation and quality degradation that are common in underwater images mentioned in the background art.

[0006] To achieve the above objectives, this invention provides a referenceless underwater image enhancement method based on morphological intelligence and sparse representation, comprising the following steps: S1. Adaptive Nonlinear White Balance: Employs a data-driven nonlinear white balance correction method to dynamically adjust the global gain and correct color shift caused by non-uniform illumination. S2. Enhanced Morphological Contrast: Employing channel-level processing methods, combining top-hat and bottom-hat transformations, enhances global contrast, highlighting hidden targets while preserving macroscopic structure; S3. Sparse representation detail enhancement: An overcomplete dictionary is constructed using non-subsampled contour wave NSCT, and cross-scale consistency features are utilized to extract fine targets and suppress scattering noise.

[0007] Preferably, the specific steps of S1 are as follows: S11. Construct a nonlinear mapping function from the input color space to the target color space to simulate and reverse the nonlinear color transformation process in the camera imaging pipeline; S12. Use a global chromaticity adjustment method to optimize the chromaticity distribution characteristics of the corrected image.

[0008] Preferably, the specific steps of S11 are as follows: S111. Input the original image and use the kernel mapping function to project the RGB color space of the original image to a high-dimensional polynomial space. S112. Calculate the RGB-uv logarithmic space histogram of the original image and encode the histogram to obtain compact features; S113. Using a pre-trained mapping function, multiple similar samples are retrieved based on compact features. The correction matrices corresponding to each sample are fused by Gaussian weighting to obtain a nonlinear correction matrix. S114. Using a reconstruction function, the nonlinear correction matrix is ​​reconstructed into a standard 3×11 matrix to complete the nonlinear white balance correction of the image and suppress the image color deviation.

[0009] Preferably, the specific steps of S12 are as follows: S121. Based on the image after nonlinear white balance correction, calculate the quantile statistics of the intensity distribution for the R, G, and B channels, and use the truncated quantiles to determine the effective dynamic range. S122. Define a global linear stretching function to stretch each channel individually, linearly mapping the dynamic range of each channel to the range [0,1]. S123. Recombine the stretched three channels to obtain a coarsely enhanced image.

[0010] Preferably, the specific steps of S2 are as follows: S21. Convert the coarse-enhanced image from the RGB color space to the HSV color space, extract the V channel components, and normalize them. interval; S22. Use top hat transformation to extract bright details and terrain transformation to extract dark details; S23. Based on the similarity of human eyes and the required spatial consistency, select the structural elements of the disk structure; S24. An optimization strategy based on local contrast enhancement rate is adopted to adaptively determine the optimal radius of the structuring element; S25. By using the optimal structural element, the difference between the bright and dark details is superimposed onto the original luminance channel to enhance contrast. S26. The optimal structuring element parameters are applied to each RGB channel, and morphological enhancement is performed independently on each channel to improve contrast and effectively preserve edge structure.

[0011] Preferably, the specific steps of S24 are as follows: S241. Define the local contrast change at each location in the image and use a dual-window model to simulate human eye perception; S242. Calculate the local contrast map of the brightness map of the coarsely enhanced image and the brightness map after morphological enhancement, and determine the contrast enhancement rate corresponding to the radius of the structuring element. S243. Through iterative search, the radius that is maximized is selected as the optimal radius.

[0012] Preferably, the specific steps of S3 are as follows: S31. Construct a basis function with multi-scale and multi-directional characteristics by using a non-subsampled pyramid and a non-subsampled directional filter bank. S32. An adaptive gain strategy based on local energy is adopted to perform nonlinear processing on the transform domain coefficients.

[0013] Preferably, the specific steps of S32 are as follows: S321. Calculate the local energy of the transform domain coefficients at each location to characterize the saliency of the image at multiple scales and in multiple directions; S322. Based on local energy, classify image regions and assign corresponding gain factors to different image regions; S323. Introduce cross-scale consistency constraints to improve the sparsity of transform domain coefficients and calculate local variance and suppression coefficient; S324. Construct a noise mask and perform soft thresholding on the transform domain coefficients by combining local variance and suppression coefficient. S325. Perform NSCT inverse transform on the transform domain coefficients after soft thresholding to obtain the reconstructed image.

[0014] Therefore, the above-mentioned reference-free underwater image enhancement method based on morphological intelligence and sparse representation has the following beneficial effects: (1) Strong scene adaptability, excellent robustness and generalization ability. This invention can be adapted to a variety of complex water quality environments and can complete stable and natural color reproduction under different water quality conditions. It effectively overcomes the defects of insufficient color fidelity and poor scene adaptability of traditional methods and has stronger stability and versatility in the face of diverse underwater scenes.

[0015] (2) Accurate color correction avoids image distortion problems. This invention has excellent color restoration capabilities. It can effectively correct color deviations in scenes with severe color distortion, and accurately restore true colors in special scenes with high contrast. It will not produce negative problems such as color difference, false color, or artifacts throughout the process, thus ensuring the authenticity and accuracy of image colors from the source.

[0016] (3) Excellent detail processing and outstanding edge enhancement effect. The present invention provides balanced and high-quality restoration effects for different areas of the image. It can achieve high-quality detail optimization and edge enhancement processing for different scene areas such as the upper scene, middle scene and lower scene of the image. At the same time, it can completely preserve the rich edge details and texture information of the image, restore the original image details to the greatest extent, and avoid the loss of details during the restoration process.

[0017] (4) Excellent imaging quality and good subjective visual effect. During the restoration process, the present invention can significantly improve the color contrast and color saturation of the image, and simultaneously optimize the image clarity and detail sharpness, making the restored image more vivid and the picture more natural, greatly improving the overall visual experience and demonstrating excellent imaging quality in subjective visual evaluation.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the framework of an embodiment of the reference-free underwater image enhancement method based on morphological intelligence and sparse representation of the present invention. Figure 2 This is a flowchart illustrating the optimal radius adaptive selection process in an embodiment of the reference-free underwater image enhancement method based on morphological intelligence and sparse representation of the present invention. Figure 3This document presents a performance comparison of different methods from the proposed no-reference underwater image enhancement method based on morphological intelligence and sparse representation on the UCCS dataset. (a) is the raw image; (b) is DICAM (Deep Inception and Channel-wise Attention Module), (c) is HFM (Hybrid Fusion Method), (d) is HF (Histformer), (e) is HUPE (Heuristic Underwater Perceptual Enhancement), (f) is PF (Phaseformer), and (g) is PCDE (Piecewise Color Correction and Dual Prior Optimized Contrast). Enhancement, segmented color correction and dual prior optimization contrast enhancement), (h) is SSUIE (spatial-spectral underwater image enhancement), (i) is U-shape (U-shape transformer), (j) is MISR (this invention). Figure 4 This is a performance comparison of different methods of the reference-free underwater image enhancement method based on morphological intelligence and sparse representation of the present invention on the ColorCheck7 dataset. (a) is Raw, (b) is DICAM, (c) is HFM, (d) is HF, (e) is HUPE, (f) is PF, (g) is PCDE, (h) is SSUIE, (i) is U-shape, and (j) is MISR. Figure 5 This is an embodiment of the reference-free underwater image enhancement method based on morphological intelligence and sparse representation of the present invention. Figure 4 Five methods that perform well in ColorCheck7 for magnified view: (a) DICAM, (b) HFM, (c) SSUIE, (d) U-shape, and (e) MISR. Figure 6This is a performance comparison of different methods of the reference-free underwater image enhancement method based on morphological intelligence and sparse representation of the present invention on the OceanDark dataset. (a) is Raw, (b) is DICAM, (c) is HFM, (d) is HF, (e) is HUPE, (f) is PF, (g) is PCDE, (h) is SSUIE, (i) is U-shape, and (j) is MISR. Figure 7 This is a performance comparison of different methods of the reference-free underwater image enhancement method based on morphological intelligence and sparse representation of the present invention on the UIEB dataset. (a) is Raw, (b) is DICAM, (c) is HFM, (d) is HF, (e) is HUPE, (f) is PF, (g) is PCDE, (h) is SSUIE, (i) is U-shape, and (j) is MISR. Figure 8 This paper presents an ablation experiment study on three candidate components in four datasets (top UCCS, ColorCheck7, OceanDark, and UIEB) for an embodiment of the no-reference underwater image enhancement method based on morphological intelligence and sparse representation of the present invention. (a) is Input, (b) is -w ACC, (c) is -w MCE, and (d) is -w SRE. Figure 9 The above is a schematic diagram of the dehazing effect of a real scene image in an embodiment of the reference-free underwater image enhancement method based on morphological intelligence and sparse representation of the present invention. (a) is a natural close-up view, (b) is a natural long-distance view, and (c) is a long-distance view of the city. Figure 10 The key point detection results are shown in the embodiment of the referenceless underwater image enhancement method based on morphological intelligence and sparse representation of the present invention. (a) is UIEB, (b) is UCCS, and (c) is OceanDark. Figure 11 The images show the edge detection results of an embodiment of the referenceless underwater image enhancement method based on morphological intelligence and sparse representation of the present invention. (a) shows the edge detection results of the original image, and (b) shows the edge detection results of the corresponding enhanced image. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0021] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning understood by one of ordinary skill in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0022] Example Please see Figures 1-11 This invention provides a referenceless underwater image enhancement method based on morphological intelligence and sparse representation, such as... Figure 1 As shown, the entire workflow of this invention consists of three adaptive core modules: ACC (Adaptive Color Consistency), MCE (Morphological Contrast Enhancement), and SRE (Sparse Representation Enhancement). These modules work collaboratively in a cascaded manner to provide a systematic solution to the three key problems in underwater image processing—color distortion, low contrast, and detail degradation—ultimately outputting image results with significantly improved visual quality. The process includes the following steps: S1, Nonlinear White Balance.

[0023] Color distortion in underwater images primarily stems from two aspects: first, the spectral imbalance caused by the selective absorption and scattering of different wavelengths of light by water; and second, the nonlinear color rendering applied by the white balance module in the camera's imaging pipeline after incorrectly estimating the light source color. Traditional linear white balance methods based on diagonal matrices cannot effectively reverse the complex nonlinear transformations in the camera's rendering pipeline. To address this issue, this invention employs a data-driven nonlinear white balance correction method. This method effectively simulates and reverses the nonlinear color transformation process in the camera's imaging pipeline by constructing a nonlinear mapping function from the input color space to the target color space. Given an input image... Its nonlinear white balance correction process can be expressed as: (1) in, The kernel mapping function projects the original RGB color space to a higher-dimensional polynomial space: (2) function This represents the vectorized polynomial correction matrix. Reconstructed into a standard 3×11 matrix form: (3) Input of function H It is a compact feature obtained by calculating and encoding the RGB-UV logarithmic space histogram of the image. In practical applications, the nearest neighbor strategy is not used to retrieve images from the training set that are similar to the current image. The K most similar samples are selected. The correction matrices corresponding to these samples are fused using Gaussian weighting to obtain the final correction matrix suitable for the current image. To achieve efficient correction, this invention directly employs a pre-trained correction function. .

[0024] Equation (3) establishes the estimation vector from the color offset. To nonlinear correction matrix After the aforementioned nonlinear white balance correction, the color deviation of the image is effectively suppressed by the direct mapping. However, the overall chromaticity ratio of the image may still be unbalanced due to excessive attenuation of the underwater red channel, resulting in an image that appears either too dark or too white. To enhance visual perception quality, this invention proposes a global chromaticity balancing method to optimize the dynamic range of each color channel. Therefore, this invention proposes a global chromaticity adjustment method aimed at optimizing the chromaticity distribution characteristics of the corrected image.

[0025] set up This represents the corrected image, for each channel. Calculate the quantile statistics of its intensity distribution. To avoid the influence of noise, truncated quantiles are used to determine the effective dynamic range: (4) in, express Quantiles and These are used to remove noise in dark areas and clipped highlights, respectively. A global linear stretching function is defined as follows: (5) Apply the above function independently to each color channel: (6) This operation linearly maps the dynamic range of each channel to the range [0,1], stretching the hue reduction state caused by the former, thereby improving the overall hue ratio. Finally, the three stretched channels are recombined to obtain a coarsely enhanced image: (7) S2, Enhanced mathematical morphological contrast.

[0026] Contrast enhancement essentially involves enhancing the dynamic range distribution of image pixel values, i.e., increasing the intensity of high-brightness areas while suppressing or reducing the grayscale values ​​of low-brightness areas, thereby expanding the overall contrast of the image. Enhancement of visual gradients can strengthen the edge and texture information of the image, improving visual clarity. This invention further extends the problem to the processing mechanism of image detail structures through set operations in mathematical morphology (such as dilation, erosion, and their combinations). Morphological operations interact with the image through structuring elements, enhancing or suppressing specific structures while preserving the main shape, thus achieving detail enhancement and noise suppression in the spatial domain.

[0027] To effectively improve image visual clarity, the core objective of this module is to adaptively enhance local contrast while maintaining overall brightness balance. Its innovation lies in the first systematic introduction of mathematical morphology methods into the field of underwater image enhancement. Through morphological top-hat and bottom-hat transformations, it extracts and processes details in the bright and dark areas of the image, thereby achieving structure-aware contrast optimization in the spatial domain.

[0028] To avoid color distortion, the image after color cast correction is first... Convert from RGB color space to HSV color space and extract its V channel components. and normalize it to The interval serves as the basis for subsequent morphological processing. This method uses top-hat transformation to extract highlight details and terrain transformation to extract shadow details. (8) in, Indicates the use of structural elements For images Perform an opening operation. Indicates the use of structural elements For images The closing operation is performed, and the structural element is selected based on human visual similarity and the required spatial consistency effect, choosing a disk-shaped structural element: (9) Here, r is the radius of the structuring element, and the choice of the radius of the structuring element will affect the overall correction effect. For example, a smaller radius can only extract very fine features, which can make up for details and is suitable for enhancing areas with rich textures, but the disadvantage is that it will ineffectively amplify noise; a larger radius enhances macroscopic contrast, but the disadvantage is that it loses important details. Therefore, the choice of r is a matter of necessity.

[0029] To adaptively determine the optimal radius This module proposes an optimization strategy based on local contrast enhancement rate (CIR). First, the local contrast change at each location in the image is defined. A dual-window model is used to simulate human eye perception: (10) in, Indicated by The mean of the inner window centered on the center. This represents the mean of the outer window. It represents a small constant that is greater than zero.

[0030] Calculate separately Local comparison of the brightness map and the brightness map after morphological enhancement. and Then the contrast enhancement rate corresponding to r is: (11) Through iterative search, the radius that is maximized is selected as the optimal parameter: (12) By utilizing the optimal structuring element, the difference between highlight and shadow details is superimposed onto the original luminance channel to enhance contrast. (13) After brightness correction, the overall tone of the image is set, and the visual effect is further enhanced. To maintain color harmony, the optimal structural element parameters are applied to each RGB channel. Morphological enhancement is performed independently for each channel: (14) in, Represents based on structural elements Morphological contrast enhancement operation. The resulting image. While enhancing contrast, it effectively preserves edge structure, laying a clear foundation for subsequent detail enhancement modules.

[0031] S3, enhanced details.

[0032] After mathematical morphological contrast enhancement, the global dynamic range and local contrast of an image are significantly improved. However, contrast enhancement is essentially a redistribution of gray values ​​in a global or local region. This process may inadvertently weaken or smooth subtle edge and fine texture information in the image while enhancing macroscopic contrast. Specifically, the enhanced image may lose some high-frequency details, resulting in blurred contours or weakened texture features, affecting visual sharpness and structural integrity. To compensate for this deficiency and further enhance the perceptual quality of the image, this stage focuses on transforming the sharpness enhancement, i.e., detail enhancement problem, into a sparse representation problem.

[0033] Within the framework of sparse representation theory, a signal can be reconstructed by selecting a linear combination of a small number of atoms from a dictionary. For example, a signal It can be represented as a dictionary matrix. With sparse coefficient vector A linear combination, i.e. The dictionary determines the efficiency of signal decomposition. To accommodate the complex ensemble structure of underwater images (such as anisotropic edges and contours), this invention utilizes non-subsampled contour wave (NSCT) to construct an overcomplete dictionary. This dictionary possesses translation invariance and strong anisotropy, enabling it to perform excellently in efficiently and sparsely representing the geometry of underwater images, while treating noise as a non-coefficient component.

[0034] Traditional orthogonal transforms (such as DCT or wavelets) are limited by the unidirectional nature of basis functions, making it difficult to sparsely represent edge information in images. NSCT, however, constructs a set of basis functions with multi-scale and multi-directional characteristics through a non-subsampled pyramid and a non-subsampled directional filter bank. Its decomposition process can be expressed as: (15) in, This represents the NSCT analysis operator (i.e., dictionary projection process). This represents the set of coefficients in the transform domain. In sparse representation, the ideal coefficients... Most coefficients should be zero (or close to zero), with only a small number of large coefficients representing the true signal. However, noise caused by water degradation also produces coefficients. Therefore, this invention requires nonlinear processing (i.e., sparse coding) of the coefficients to distinguish between "true details" and noise. This invention proposes an adaptive gain strategy based on local energy, where local energy reflects the significance of the location coefficient across multiple scales and directions. The calculation formula is as follows: (16) in, Representing the m Layer scale, first n Transformation coefficients in each direction. Based on This invention divides image regions into different categories and applies different gain factors. : (17) In this invention, the E-cap is used to represent local energy. , (i=1234) represent strong structure regions, medium structure regions, weak structure regions, and flat / noisy regions, respectively. represent true edges, weakly combined details and noise, noise drowning out minute details, and background or pure noise, respectively.

[0035] To further improve sparsity, this invention introduces a cross-scale consistency constraint. Real-world geometry typically exhibits responses at both coarse and fine scales, while noise is usually random and isolated.

[0036] (18) (19) in, For local variance, For the inhibition factor, η∈[0,1].

[0037] This invention constructs a noise mask Combining the strategies in formulas (17) and (18), soft thresholding was applied to the coefficients. Finally, the image was reconstructed using the inverse transform (synthesis operator) of NSCT: (20) Through this sparse representation framework based on an overcomplete dictionary, the method of this invention is mathematically equivalent to finding the optimal coefficient approximation solution for underwater images under a geometric dictionary. This not only achieves accurate enhancement of smiling targets but also effectively proposes non-sparsely distributed scattering noise by utilizing sparsity priors.

[0038] S4. Experiment and Conclusion.

[0039] This invention provides a comprehensive qualitative and quantitative evaluation of the proposed no-reference underwater image enhancement method (MISR) based on morphological intelligence and sparse representation. To verify its advancement and effectiveness, the MISR algorithm is compared with eight state-of-the-art (SOTA) algorithms released in the past three years. Experiments were conducted on four widely used underwater image datasets, using six no-reference image quality metrics as quantitative benchmarks. Furthermore, the completeness and practical applicability of the algorithm are verified through ablation experiments and extended application tests. All experiments were implemented using MATLAB 2021a on a mobile workstation equipped with an Intel Core i9-14900HX processor and an NVIDIA 4060 graphics card. Due to space limitations, only representative images are shown. Extensive replication experiments demonstrate that MISR outputs stable results, outperforming mainstream algorithms of the past three years in terms of visual quality, robustness, and accuracy. Subsequent ablation experiments further confirm the effectiveness of each module and the importance of their synergistic effects. Experimental parameters have been rigorously validated to ensure reliability; extended experiments in keypoint detection, object detection, and image dehazing demonstrate the scalability of the algorithm in practical applications. This invention is considered to have significant value in promoting the practical application of underwater vision technology.

[0040] To ensure a comprehensive and rigorous evaluation of the proposed MISR method, this invention selected eight representative underwater image enhancement (UIE) algorithms published in the past three years as benchmarks. These methods cover a variety of technical paradigms, from traditional physical models to cutting-edge deep learning frameworks, thus guaranteeing the breadth and depth of the evaluation. In the traditional method category, this invention selected SPL, based on a physical imaging model, and HFM, which relies on non-physical prior knowledge; these two represent classic paradigms utilizing underwater light transmission characteristics and image statistical features, respectively. For deep learning methods, this invention further divides them into two main directions: Convolutional Neural Networks (CNN) and Transformer architectures, to comprehensively examine the impact of different network structures on the enhancement effect. Specifically, CNN-based methods include: DICAM, which incorporates traditional physical models to improve interpretability; HUPE, which introduces an attention mechanism to focus on key features; and SSUIE, which employs a cross-domain fusion strategy to integrate multi-source information. For Transformer-based methods, this invention selects Histformer to improve the basic architecture; Phaseformer combined with Generative Adversarial Networks (GANs) to enhance visual realism; and U-shape Transformer fused with the U-Net structure to enhance feature reuse capabilities. This carefully selected set of eight algorithms all achieve or approach state-of-the-art (SOTA) performance levels, laying a solid foundation for verifying the advancement and robustness of the MISR method across different technical scenarios.

[0041] To construct a comprehensive and rigorous testing environment, this invention employs four of the most commonly used open-source benchmark datasets in the underwater vision field: UCCS (Underwater Color Cast Set), ColorCheck7, OceanDark (Low-Light Underwater Image Enhancement), and UIEB (Underwater Image Enhancement Set). These datasets are specifically designed to address different quality degradation problems in underwater imaging, enabling this invention to evaluate the performance limits and generalization capabilities of the algorithm from multiple dimensions.

[0042] The UCCS dataset contains images captured under three different water quality conditions, characterized by severe color cast and spectral attenuation. This dataset is primarily used to quantitatively evaluate the color correction capability and spectral restoration accuracy of the algorithm under extreme quality degradation conditions. The ColorCheck7 dataset, built upon a standard reference color chart, provides objective color benchmark values; it allows for precise testing of indicators such as color deviation and color fidelity. Addressing the challenges of underwater imaging in low-light environments, this invention specifically introduces the OceanDark dataset. This dataset contains a series of dark underwater images with complex scenes and extremely low signal-to-noise ratios (SNR); related experiments aim to verify whether the algorithm can effectively suppress noise, enhance low-light performance, preserve fine texture details, and correct color distortion. Furthermore, the UIEB dataset, renowned for its high resolution and rich color information, is primarily used to evaluate the improvement in overall visual sharpness after color restoration and to detect artifacts or over-enhancement. The combination of these two datasets provides a solid empirical foundation for verifying the robustness and adaptability of the proposed method under different image degradation mechanisms.

[0043] To ensure an objective and quantitative analysis of the performance of the proposed method and eight comparative algorithms, this invention employs a recognized no-reference image quality assessment system. This system encompasses multiple dimensions, including overall perceptual quality, structural sharpness, edge detail retention, and statistical properties, thereby comprehensively reflecting the effectiveness of the enhancement algorithms in restoring underwater image details and visual fidelity.

[0044] This invention employs six core metrics: UIQM (Underwater Image Quality Metric), UISM (Underwater Image Structure Metric), AG (Average Gradient, reflecting edge retention and detail richness), EI (Edge Intensity, characterizing image edge energy levels), UCIQE (Underwater Color Image Quality Evaluation Metric, comprehensively reflecting contrast, chroma, and saturation), and VAR (Variance, as a simplified alternative to image entropy, characterizing the overall variation characteristics of an image). Among these, UIQM, as a comprehensive perceptual quality metric, reflects the overall visual effect from multiple perspectives; UISM and AG focus on evaluating image sharpness and texture detail richness, their values ​​directly reflecting the algorithm's ability to retain high-frequency information; EI and VAR are used to measure image sharpness and contrast; and UCIQE is specifically used to evaluate color richness and contrast. It is important to note that all the above metrics follow a unified evaluation standard: higher values ​​indicate better performance in the corresponding dimensions (such as visual quality, sharpness, and edge detail retention). Through this multi-dimensional quantitative evaluation framework, this invention enables precise horizontal comparison and in-depth analysis of the performance improvement effects of different algorithms.

[0045] S41, Qualitative Experiment.

[0046] like Figure 3 The image shows a performance comparison of different methods on the UCCS dataset. For clarity, one example was randomly selected from each of the three subsets of UCCS. Here, HF represents Histformer, PF represents Phaseformer, and U-shape represents U-shaped Transformer.

[0047] Figure 3This paper presents qualitative comparison results of randomly selected samples from the UCCS dataset. Divided into three subsets (blue, blue-green, and green) based on water quality characteristics, the aim is to systematically evaluate the color restoration capabilities of each algorithm under different spectral attenuation conditions. In the first subset, the original images exhibit typical underwater imaging characteristics: the blue channel intensity is significantly higher than the green channel. Analysis shows that while method (c) HFM effectively corrects the color of the subject, overcorrection occurs in the background area, resulting in an unnatural purple hue. Method (g) SPL performs poorly in restoring both the subject and background, failing to effectively improve image quality. The other comparative methods, while restoring the subject to some extent, typically darken the background area, affecting the overall visual effect. In the second subset, the original images present a complex degraded scene with high values ​​in both the blue and green channels. Visually, methods (b) DICAM, (c) HFM, (h) SSUIE, and (i) U-shaped transformer perform excellently in color correction, accurately restoring the true tones of the scene. In contrast, the image processed by (g) SPL shows a significant green enhancement, while other methods typically exhibit a yellowish tint, indicating that their color fidelity needs improvement. In the third subset, the original images have low values ​​for both the blue and green channels, posing a significant challenge to the algorithm's low-light enhancement capabilities. Visual analysis shows that methods (b) DICAM, (c) HFM, and (h) SSUIE all maintain stable color correction effects; while images processed by (e) HUPE and (g) SPL show varying degrees of green shift, and other comparative methods also exhibit significant yellow bias. In contrast, the proposed method (j) MISR achieves stable and natural color reproduction under all three water quality conditions, effectively overcoming the limitations of traditional methods in terms of color fidelity and scene adaptability, demonstrating excellent robustness and generalization ability.

[0048] like Figure 4 The diagram shows a performance comparison of different methods on the ColorCheck7 dataset. Three samples from ColorCheck7 are randomly selected for this illustration. HF represents Histformer, PF represents Phaseformer, and U-shape represents U-shaped Transformer.

[0049] Figure 4This paper presents the results of color correction on the ColorCheck7 dataset using a color chart, aiming to further verify the subjective correctness and accuracy of the color restoration algorithm. To evaluate the performance of each method in more detail, this invention defines the color block in the upper left corner of the color chart as the origin (0,0). Subsequently, this invention conducts independent magnified comparative analysis on four contrasting methods (b) DICAM, (c) HFM, (h) SSUIE, and (i) U-shaped transformer that demonstrate excellent visual effects in the UCCS dataset.

[0050] like Figure 5 The second row of images shows that the acquired color chart area suffers from some degree of image degradation due to interference from water ripples during the original shooting process. Under these complex conditions, the proposed (j) MISR method demonstrates significantly higher accuracy and naturalness in color restoration compared to the other four algorithms, and is also more effective in suppressing environmental interference. The third row of images clearly shows that the method of this invention can effectively distinguish color blocks with similar hues (such as peach pink (coordinate 2,1) and standard red (coordinate 2,2)), demonstrating its significant advantages in identifying subtle color differences and maintaining color sharpness. However, in the fourth row, there are significant differences in the restoration results of all the comparison methods. For special color blocks with low light and low saturation (such as indigo and jet black), the existing algorithms are not performing well. This indicates that current underwater image enhancement technology still faces challenges in processing color details in extremely dark areas. Nevertheless, the method of this invention still has significant advantages: its core innovation lies in significantly improving the color contrast and saturation of the restored image while ensuring higher clarity and detail sharpness, thus presenting excellent image quality in subjective visual evaluation.

[0051] like Figure 6 The image shows a performance comparison of different methods on the OceanDark dataset. To illustrate this, three examples were randomly selected from OceanDark. Here, HF represents Histformer, PF represents Phaseformer, and U-shape represents U-shaped Transformer.

[0052] Figure 6Qualitative test results of detail recovery capabilities under different scenarios are presented. Through visual comparative analysis of several typical underwater scenes in this image, this invention can deeply evaluate the performance differences of various algorithms in handling complex textures and edge preservation. First, in the middle scene (the main subject is an underwater operating robot covered with a large amount of algae), it can be observed that (b) the DICAM method has obvious overexposure, resulting in the loss of texture details on the algae surface and blurry object outlines; second, in the upper scene containing seashells and pebbles, although (g) the SPL method effectively improves the overall contrast and makes some details more clearly visible, it still has limitations in handling small objects; in addition, all the comparison algorithm results show significant color distortion, with the overall tone presenting an unnatural purple hue, which seriously damages the realism of the scene.

[0053] In comparison, the (j) MISR method proposed in this invention demonstrates superior performance in detail processing and edge enhancement. Specifically, in the upper scene of a flat seabed area, the method not only accurately restores the color of the shells but also achieves a refined restoration effect. Regarding refined peripheral detail processing: in the middle scene, by adaptively enhancing the contrast of subtle object edges, the outline of the robot covered with algae exhibits excellent visual clarity, effectively avoiding detail blurring caused by overexposure; in the lower scene, the method successfully solves the color cast problem—not only highlighting the main features of swimming fish but also making microstructures such as buttons on the brown panel in the background clearly visible. This result fully demonstrates the advantages of the (j) MISR method in balancing color correction and detail restoration.

[0054] like Figure 7 The image shows a performance comparison of different methods on the UIEB dataset. For ease of explanation, three examples were randomly selected from the UIEB dataset. Here, HF represents Histformer, PF represents Phaseformer, and U-shape represents U-shaped Transformer.

[0055] exist Figure 7 In this study, the main focus was on evaluating the sharpness and detail reconstruction capabilities of various algorithms on high-resolution images. In the coral reef scene above, visual observation showed that all contrast algorithms experienced varying degrees of detail loss during the enhancement process, resulting in blurred texture information and degraded image quality. Given the potential for subjective bias in macroscopic visual judgment, this invention further selected a scene containing swimming fish in the middle for microscopic analysis.

[0056] In this scenario, it is evident that method (c) HFM introduces false color distortion and causes significant overexposure in the lower right fish area, resulting in the complete loss of local details. Methods (b) DICAM, (d) Histformer, (f) Phaseformer, and (i) U-shape generally suffer from blurred edges and loss of high-frequency details, making the fish outline unclear and texture features difficult to discern. In contrast, the proposed method (j) MISR significantly outperforms the aforementioned algorithms in terms of visual clarity—it not only achieves accurate color reproduction but also effectively preserves rich edge details and texture information, presenting a more natural and vivid visual effect.

[0057] S42, Quantitative Experiment.

[0058] As shown in Table 1, this invention presents the average values ​​of underwater image enhancement evaluation metrics for MISR and eight other advanced algorithms on three different quality degradation subsets of the UCCS dataset. By calculating the average values ​​of UIQM, UISM, AG, EI, and UCIQE, the evaluation results of degraded underwater images can be converted into numerical values ​​for quantitative comparison: the higher the values ​​of the above metrics, the better the enhancement effect. As shown in Table 1, MISR performs excellently in color restoration and edge detail preservation in underwater image enhancement; however, there is still room for improvement in contrast and saturation of deep-degraded images. The results in Table 1 fully demonstrate the significant advantages of MISR in image color balance and provide strong evidence for the effectiveness of this invention.

[0059] Table 1. Average evaluation metrics of underwater image enhancement using MISR and eight other advanced algorithms on three different quality degradation subsets of the UCCS dataset.

[0060] Note: Bold data is optimal, and underlined data is suboptimal.

[0061] As shown in Table 2 below, this invention presents the average values ​​of underwater image enhancement evaluation metrics for MISR and eight other advanced algorithms on three different quality degradation subsets of the ColorCheck7 dataset. By calculating the average values ​​of UIQM, UISM, AG, EI, and UCIQE, the evaluation results for degraded underwater images are converted into numerical forms for quantitative comparison: higher values ​​indicate more significant enhancement effects. As shown in Table 2, MISR performs best in color reproduction across all color types; MISR achieves optimal results in both visual quality and color reproduction. Its significant advantages in numerical evaluation further confirm the outstanding contributions of the algorithm of this invention to the field of color reproduction.

[0062] Table 2 shows the average evaluation metrics of underwater image enhancement for MISR and eight other advanced algorithms on three different quality degradation subsets of the ColorCheck7 dataset.

[0063] Note: Bold data is optimal, and underlined data is suboptimal.

[0064] As shown in Table 3, this invention presents the average values ​​of underwater image enhancement evaluation metrics for the MISR algorithm and eight other cutting-edge algorithms on three different quality-degraded subsets of the OceanDark dataset. By calculating the average values ​​of UIQM, UISM, AG, EI, and UCIQE, the evaluation results of degraded underwater images can be converted into numerical values ​​for quantitative comparison. Higher values ​​for these metrics indicate better enhancement performance. This dataset is specifically designed to measure the detail recovery capability of this algorithm, focusing on AG (edge ​​preservation), UISM (detail texture recovery), and EI (edge ​​gradient strength) to verify the algorithm's advantages. As shown in the table, the algorithm of this invention exhibits significant advantages, and these results provide reliable evidence for the algorithm's detail enhancement capabilities.

[0065] Table 3. Average evaluation metrics of underwater image enhancement using the MISR algorithm and eight other cutting-edge algorithms on three different quality degradation subsets of the OceanDark dataset.

[0066] Note: Bold data is optimal, and underlined data is suboptimal.

[0067] As shown in Table 4, this invention presents the average values ​​of underwater image enhancement evaluation metrics for the MISR algorithm and eight other cutting-edge algorithms on three different quality degradation subsets of the UIEB dataset. By calculating the average values ​​of UIQM, UISM, AG, EI, and UCIQE, the evaluation results of degraded underwater images are converted into numerical form for quantitative comparison. Higher values ​​for these metrics indicate superior enhancement effects. As demonstrated by the excellent results shown in Table 4, the algorithm of this invention exhibits strong competitiveness in both high-resolution and high-contrast blurred scenes. The quantitative results above show that this algorithm can not only effectively correct colors in scenes with severe color distortion, but also accurately restore colors without producing color difference or false color artifacts under high contrast conditions. Improving image sharpness while ensuring detail and edge clarity is the core advantage of this algorithm compared to other algorithms.

[0068] Table 4. Average evaluation metrics of underwater image enhancement using the MISR algorithm and eight other cutting-edge algorithms on three different quality degradation subsets of the UIEB dataset.

[0069] Note: Bold data is optimal, and underlined data is suboptimal.

[0070] S43, Ablation Experiment.

[0071] In this invention, systematic and comprehensive ablation experiments were conducted on four representative and challenging underwater image benchmark datasets (UCCS, ColorCheck7, OceanDark, and UIEB). The core objective of these experiments was to rigorously evaluate the effectiveness of each core module in the proposed algorithm from both quantitative and qualitative dimensions. Specifically, the focus was on three key modules: the Adaptive Color Consistency (ACC) module for color correction, the Mathematical Contrast Enhancement (MCE) module for contrast enhancement, and the Sparse Representation Enhancement (SRE) module for detail and texture preservation and enhancement.

[0072] like Figure 8 As shown, the image processing effects of each module being introduced step by step are visually compared, clearly revealing the functional differences of each algorithm component under different underwater imaging environments. The original input images (leftmost column) generally exhibit typical color cast phenomena caused by the selective absorption of light wavelengths by water and forward / backward scattering effects, with the overall tone generally biased towards blue-green or yellowish-brown, severely impairing the image's visual quality. After processing by the ACC module, the extreme color cast effect is significantly suppressed, and the overall image tone tends to be more natural and realistic, fully demonstrating the module's effectiveness in restoring color constancy underwater. After introducing the MCE module (i.e., the contrast enhancement module, corresponding to the aforementioned MCE module), the overall contrast and basic visual details of the image are significantly restored: information in dark areas becomes clearly visible, and highlight areas are reasonably compressed, fully demonstrating the module's key role in improving dynamic range and local contrast. Furthermore, after integrating the SRE module into the complete framework, the high-frequency details and edge structure of the image are also significantly enhanced. For example, in the UCCS dataset, the natural texture patterns on underwater rock surfaces become more clearly discernible; in the UIEB dataset, the contour structures around the eye region of fish targets are sharper due to enhanced detail. These qualitative results demonstrate that each module has targeted processing capabilities for different image degradation features.

[0073] In the specific configuration of the ablation experiment, this invention performs systematic module pruning on the complete model to separate and evaluate the independent contribution of each component to the final enhancement effect. `-w ACC` indicates that the adaptive color consistency module is retained for color correction, while subsequent modules are removed. The contrast and detail enhancement modules can be analyzed independently for the effect of color correction. `-wMCE` indicates the introduction of a morphological contrast enhancement module to improve contrast, while removing the color correction and detail enhancement modules, used to evaluate the independent effect of this module in improving grayscale distribution and local dynamic range; `-w SRE` indicates the introduction of a sparse representation enhancement module to achieve multi-scale, multi-directional high-order detail texture enhancement, and the removal of the first two modules to evaluate the contribution of transform domain processing alone to texture recovery. Each pruned version and the complete model are evaluated on the same test dataset to ensure the fairness and consistency of the ablation experiment results.

[0074] The experimental results present clearly defined physical metrics: UIQM, UISM, AG, EI, UCIQE, and VAR. By comparing the performance of each pruned version with the complete model on these metrics, this invention can systematically verify the marginal contribution of each module to the final enhancement effect.

[0075] Table 5 Ablation Experiment Results Based on UCCS Dataset

[0076] Table 6 Ablation experiment results based on ColorCheck7, Oceanark, and UIEB datasets

[0077] Note: The bolded data in Tables 5 and 6 represents the optimal values.

[0078] Based on the quantitative results in Tables 5 and 6, the following conclusions can be drawn: Each module is functionally independent, with no redundancy or conflicting effects; simultaneously, when working together, they produce significant synergistic gains, forming a complete and robust enhancement framework. The changing trend of the UCIQE index clearly shows that the most significant improvement in image contrast and saturation is achieved after introducing the MCE module alone, with its UCIQE value even exceeding that of versions using only the ACC module or only the SRE module. This fully demonstrates that the MCE module can effectively compensate for the local saturation decrease that may be caused by the ACC module during color balancing and make up for the contrast compression effect brought about by color correction. On the other hand, the enhancement effect of the SRE module on image detail and texture has been clearly and consistently verified by the AG and EI indices: specifically, in terms of the AG index, the model variant containing the SRE module significantly outperforms the version without this module, reflecting its unique advantages in high-frequency component reconstruction and texture structure preservation. From the visual comparison results, the proposed method not only effectively improves image sharpness and color reproduction but also significantly suppresses the generation of false colors and halo artifacts, exhibiting excellent visual naturalness and structural integrity. When the three modules are fully integrated and work together, the core advantages of each module are fully utilized and complement each other, thereby achieving the best underwater image enhancement effect in both qualitative and quantitative dimensions.

[0079] S44, Extended Applications.

[0080] The core objective of underwater image enhancement technology is to support subsequent tasks in machine vision and complex, advanced applications. To further verify the effectiveness of the algorithm in this invention, its performance was evaluated in three tasks: real-world image dehazing, keypoint detection, and underwater object detection.

[0081] The image dehazing experiment aims to verify the algorithm's ability to improve image sharpness. For example... Figure 9 Experimental results show that the unprocessed original image exhibits obvious blurring and color shift, while the image processed by this algorithm shows significant detail enhancement. The method of this invention effectively suppresses noise interference and significantly expands the contrast difference between pixels, thereby significantly improving the overall image clarity. When evaluating dehazing performance, this invention uses subjective visual sharpness as the core evaluation index to meet the most basic visual perception requirements in the field of image enhancement.

[0082] To more quantitatively evaluate the image detail recovery capability, this invention conducted key point detection experiments, such as... Figure 10As shown, the original severely degraded image exhibits severe edge smoothing, resulting in overall blurring and a significant reduction in the number of keypoints. In contrast, the image quality processed by this algorithm is significantly improved, directly reflected in the substantial increase in the number of keypoints. Experimental data shows that the number of keypoints increases by 2 to 150 times compared to before processing, a result that fully demonstrates the superior performance of this algorithm in image detail restoration.

[0083] Regarding edge detection experiments, such as Figure 11 As shown, this invention employs a Gaussian function for image smoothing preprocessing, which effectively suppresses noise interference while preserving key structural features. Subsequently, gradient operators (such as Sobel or Canny) are used to perform differentiation operations on the smoothed image, thereby accurately extracting edge information. Experimental results show that the edges of the image enhanced by this algorithm exhibit clearer and more continuous response features, and the edge localization accuracy is significantly improved, further verifying the effectiveness of the algorithm in enhancing image structural clarity. Through comprehensive evaluation of image dehazing, keypoint detection, and edge detection, this invention not only verifies the algorithm's superior performance in improving image clarity but also demonstrates its wide applicability and enormous practical application potential. These experimental results not only provide strong support for the research of this invention but also lay a solid foundation for subsequent exploration in the field of machine vision.

[0084] Therefore, the above-mentioned referenceless underwater image enhancement method based on morphological intelligence and sparse representation can accurately distinguish noise from real details, achieve excellent clarity and texture restoration effect, and has effectiveness and superiority in image dehazing, key point detection and edge detection.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A reference-free underwater image enhancement method based on morphological intelligence and sparse representation, characterized in that, Includes the following steps: S1. Adaptive Nonlinear White Balance: Employs a data-driven nonlinear white balance correction method to dynamically adjust the global gain and correct color shift caused by non-uniform illumination. S2. Enhanced Morphological Contrast: Employing channel-level processing methods, combining top-hat and bottom-hat transformations, enhances global contrast, highlighting hidden targets while preserving macroscopic structure; S3. Sparse representation detail enhancement: An overcomplete dictionary is constructed using non-subsampled contour wave NSCT, and cross-scale consistency features are utilized to extract fine targets and suppress scattering noise.

2. The reference-free underwater image enhancement method based on morphological intelligence and sparse representation according to claim 1, characterized in that, The specific steps of S1 are as follows: S11. Construct a nonlinear mapping function from the input color space to the target color space to simulate and reverse the nonlinear color transformation process in the camera imaging pipeline; S12. Use a global chromaticity adjustment method to optimize the chromaticity distribution characteristics of the corrected image.

3. The reference-free underwater image enhancement method based on morphological intelligence and sparse representation according to claim 2, characterized in that, The specific steps of S11 are as follows: S111. Input the original image and use the kernel mapping function to project the RGB color space of the original image to a high-dimensional polynomial space. S112. Calculate the RGB-uv logarithmic space histogram of the original image and encode the histogram to obtain compact features; S113. Using a pre-trained mapping function, multiple similar samples are retrieved based on compact features. The correction matrices corresponding to each sample are fused by Gaussian weighting to obtain a nonlinear correction matrix. S114. Using a reconstruction function, the nonlinear correction matrix is ​​reconstructed into a standard 3×11 matrix to complete the nonlinear white balance correction of the image and suppress the image color deviation.

4. The reference-free underwater image enhancement method based on morphological intelligence and sparse representation according to claim 3, characterized in that, The specific steps of S12 are as follows: S121. Based on the image after nonlinear white balance correction, calculate the quantile statistics of the intensity distribution for the R, G, and B channels, and use the truncated quantiles to determine the effective dynamic range. S122. Define a global linear stretching function to stretch each channel individually, linearly mapping the dynamic range of each channel to the range [0,1]. S123. Recombine the stretched three channels to obtain a coarsely enhanced image.

5. The reference-free underwater image enhancement method based on morphological intelligence and sparse representation according to claim 4, characterized in that, The specific steps of S2 are as follows: S21. Convert the coarse-enhanced image from the RGB color space to the HSV color space, extract the V channel components, and normalize them. interval; S22. Use top hat transformation to extract bright details and terrain transformation to extract dark details; S23. Based on the similarity of human eyes and the required spatial consistency, select the structural elements of the disk structure; S24. An optimization strategy based on local contrast enhancement rate is adopted to adaptively determine the optimal radius of the structuring element; S25. By using the optimal structural element, the difference between the bright and dark details is superimposed onto the original luminance channel to enhance contrast. S26. The optimal structuring element parameters are applied to each RGB channel, and morphological enhancement is performed independently on each channel to improve contrast and effectively preserve edge structure.

6. The referenceless underwater image enhancement method based on morphological intelligence and sparse representation according to claim 5, characterized in that, The specific steps of S24 are as follows: S241. Define the local contrast change at each location in the image and use a dual-window model to simulate human eye perception; S242. Calculate the local contrast map of the brightness map of the coarsely enhanced image and the brightness map after morphological enhancement, and determine the contrast enhancement rate corresponding to the radius of the structuring element. S243. Through iterative search, the radius that is maximized is selected as the optimal radius.

7. The reference-free underwater image enhancement method based on morphological intelligence and sparse representation according to claim 5, characterized in that, The specific steps of S3 are as follows: S31. Construct a basis function with multi-scale and multi-directional characteristics by using a non-subsampled pyramid and a non-subsampled directional filter bank. S32. An adaptive gain strategy based on local energy is adopted to perform nonlinear processing on the transform domain coefficients.

8. The reference-free underwater image enhancement method based on morphological intelligence and sparse representation according to claim 7, characterized in that, The specific steps of S32 are as follows: S321. Calculate the local energy of the transform domain coefficients at each location to characterize the saliency of the image at multiple scales and in multiple directions; S322. Based on local energy, classify image regions and assign corresponding gain factors to different image regions; S323. Introduce cross-scale consistency constraints to improve the sparsity of transform domain coefficients and calculate local variance and suppression coefficient; S324. Construct a noise mask and perform soft thresholding on the transform domain coefficients by combining local variance and suppression coefficient. S325. Perform NSCT inverse transform on the transform domain coefficients after soft thresholding to obtain the reconstructed image.

Images