An underwater mollusk larvae image enhancement method integrating wavelet transform and Mamba architecture

Abstract

This invention proposes an underwater mollusk larvae image enhancement method integrating wavelet transform and Mamba architecture, comprising: establishing a U-Net network based on wavelet transform; in the encoding stage, constructing a filter bank based on discrete wavelet transform (DWT) to perform frequency band separation of multi-scale features; constructing a three-layer wavelet decomposition cascaded downsampling architecture, and realizing information enhancement of multi-frequency band features through an interactive dual-branch framework, wherein the interactive dual-branch framework includes a low-frequency Mamba processing module and a high-frequency window enhancement module; in the decoding stage, performing spatial domain reconstruction of multi-frequency band features using inverse wavelet transform (IDWT); and realizing visible light reconstruction of the target image based on the above-mentioned U-Net network enhancement. The model constructed by the method of this invention is used to effectively suppress feature loss caused by traditional downsampling while inheriting the advantages of U-shaped network topology through frequency band decoupling, dual-path interaction, and full information reconstruction mechanisms, thereby achieving high-fidelity restoration of image content.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to an image enhancement method for underwater mollusk larvae that integrates wavelet transform and Mamba architecture. Background Technology

[0002] Shellfish aquaculture (such as oysters, scallops, and mussels) is an important part of my country's marine economy. In the shellfish seedling cultivation process, larval density statistics and growth status monitoring are crucial steps in determining the success rate of seedling cultivation. Traditional manual microscopic counting methods are not only time-consuming and labor-intensive but also prone to significant subjective errors, making them unsuitable for large-scale industrial seedling cultivation.

[0003] Therefore, an automatic identification and counting system for underwater shellfish larvae based on computer vision has emerged. However, underwater shellfish larvae images suffer from degradation due to the absorption and scattering of light by the water and the limitations of artificial light sources. These degradations include uneven lighting, low contrast and color distortion, and severe noise interference, directly affecting the accuracy of subsequent identification algorithms. Image enhancement of shellfish larvae has become a core challenge in this field. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes an underwater mollusk larvae image enhancement method that integrates wavelet transform and Mamba architecture. This method solves the problems of insufficient feature extraction, difficulty in processing global information, susceptibility to noise and artifact interference, and poor robustness in existing underwater mollusk larvae image enhancement methods. Furthermore, this invention can effectively improve the image enhancement effect of underwater mollusk larvae, enhance image details and local contrast, and restore more accurate and natural colors.

[0005] The technical solution adopted by the present invention to achieve the above objectives is as follows:

[0006] A method for enhancing underwater mollusk larvae images by fusing wavelet transform and Mamba architecture is proposed. The method involves obtaining an image enhancement network model based on the fusion of wavelet transform and Mamba architecture. This model, through frequency band decoupling, dual-path interaction, and full-information reconstruction mechanisms, effectively suppresses feature loss caused by traditional downsampling while inheriting the advantages of U-shaped network topology, achieving high-fidelity restoration of image content. The method includes the following steps:

[0007] Establishing a U-Net network based on wavelet transform, including:

[0008] In the encoding stage, a filter bank is constructed based on Discrete Wavelet Transform (DWT) to separate the frequency bands of multi-scale features; a three-layer wavelet decomposition cascaded downsampling architecture is constructed, and information enhancement of multi-band features is achieved through an interactive dual-branch framework. The interactive dual-branch framework includes a low-frequency Mamba processing module and a high-frequency window enhancement module; wherein, the low-frequency Mamba processing module is used to handle low-frequency component color distortion and high-frequency component detail blur; the high-frequency window enhancement module is used for cross-band feature injection and frequency band reconstruction fusion.

[0009] In the decoding stage, inverse wavelet transform (IDWT) is used to reconstruct the multi-band features in the spatial domain, thereby achieving spatial domain reconstruction of the target image. Visible light reconstructed image based on the aforementioned U-Net network enhancement .

[0010] The Discrete Wavelet Transform (DWT) is performed using the following steps:

[0011] For a given target image The Haar wavelet basis functions were used for decomposition, and four sets of subband features were obtained through cascading and tensor slicing operations:

[0012]

[0013] Among them, low-frequency components Encoding global color distribution and illumination characteristics, high-frequency components It captures edge and texture details in the vertical, horizontal, and diagonal directions, respectively.

[0014] The interactive dual-branch framework of the sampling architecture for each wavelet decomposition layer includes:

[0015] In the low-frequency enhancement path, a low-frequency Mamba processing module is used for cascaded multi-scale long-range feature fusion. By modeling cross-level spatial dependencies, the fused image downsampling information is processed. The low-frequency components are enhanced and strengthened to improve information.

[0016] In the high-frequency optimization path, a high-frequency window enhancement module is used to refine the details of the fused high-frequency components under the constraint of low-frequency information.

[0017] The low-frequency Mamba processing module performs layer normalization on the input. Stable feature distribution is obtained The Mamba fusion scan block captures long-range structural dependencies between pixels through selective scanning operators, introduces random depth layers, and utilizes deep convolution and gating mechanisms of a feedforward network to enhance the spatial local interaction of features. Finally, the enhanced low-frequency features are output by a normalization layer. .

[0018] Input features After entering the Mamba fusion scan block, it is projected to the extended dimension via a linear layer and split into two branches. and ;

[0019] In the first branch, features After adjusting the dimensions, the feature channels become ,in As an expansion factor, the dimension-adjusted features are subjected to depthwise convolution to extract local features, and then... The activation introduces nonlinear features, which are then combined with a selected state-space model (SSGM) to achieve efficient spatial feature modeling through a dynamic scan path and attention fusion mechanism. Finally, the model is mapped back to the original number of channels through a linear layer. This branching process is represented as follows:

[0020]

[0021] in, Indicates the normalization layer. This indicates the selection of a state-space model. Represents a function, This represents depthwise separable convolution. Indicates a linear layer;

[0022] In the second branch, from the projection split ,go through The activation process is represented as follows:

[0023]

[0024] The two branches are finally accumulated by Hadamard, and then passed through a linear layer to produce an output of the same size as the input. The overall input-output relationship of the Mamba fusion scan block is represented as follows:

[0025]

[0026] in, This represents the Hadamard product. This indicates a linear layer.

[0027] The selected state space model includes two modules: a multi-scale feature scanning module and a shuffling attention module.

[0028] The multi-scale feature scanning module dynamically generates four parallel scanning paths based on the resolution of the input image. Each path employs four different scanning directions: forward, reverse, transposed forward, and transposed reverse. By introducing scanning length and offset length parameters, selective multi-scale feature extraction is achieved, obtaining information covering everything from local details to global structure. Then, the 2D feature map is converted into a 1D sequence based on the scanning path, and the long-distance dependencies of the sequence features are further captured.

[0029] The shuffling attention module will input features After splicing along the channel dimension, they are all divided into Groups are obtained by mixing them through dimensional recombination. This enhances cross-channel information flow; channel weights are generated by performing nonlinear transformations on the compressed features through global average pooling layers and grouped convolutions. The generated channel weights are applied to the shuffled features to selectively enhance the features, resulting in selective enhancement. The goal is to enhance key features in underwater shellfish larvae images and suppress noise and redundant information by using information exchange between channels and adaptive weight allocation.

[0030] The process of the mixed-wash attention module is represented as follows:

[0031]

[0032]

[0033]

[0034]

[0035] in, Channel splicing operation For channel mixed washing operation, For average pooling layer operations, For convolution operations, For learnable convolution kernel parameters, Use the Sigmoid activation function;

[0036] Regroup the shredded channels: ,in, Features after regrouping Regroup the channels.

[0037] The high-frequency window enhancement module first maps high-frequency features to a local window space using a window partitioning strategy, reducing the computational complexity of attention. Secondly, it introduces low-frequency features to enrich the high-frequency features and filters key details through feature matching. Specifically, it optimizes convolution operations using low-rank matrix factorization, enhancing the ability to capture effective correlations while compressing model parameters. The module includes the following steps:

[0038] First, the high-frequency components Normalized high-frequency features With enhanced low-frequency characteristics Divide the window into 8x8 sections to obtain local window features. , where i represents the i-th window, focusing on the local context and reducing redundant global calculations;

[0039] Within the window, features The query (Q), key (K), and value (V) are generated through 1×1 convolution, and local spatial associations are enhanced through 3x3 depthwise convolution. A cross-feature matching fusion unit is introduced to filter nearest neighbor matching results to optimize the feature representation of Q. Subsequently, the channel attention within the window is calculated through a multi-head self-attention mechanism.

[0040]

[0041] in, For channel attention output, For the attention mechanism, the query matrix, key matrix, and value matrix are... For function, It is a coefficient that controls the size of the Q-K dot product;

[0042] The output of the i-th window is then obtained through residual fusion. and all window features By stitching back the feature map to its original size, we obtain ,

[0043] Finally, the channel features are unified through a feedforward network containing 1x1 convolutions to obtain the final output. ;

[0044]

[0045]

[0046]

[0047] in, For cross-feature matching and fusion units; It is a feedforward network.

[0048] The cross-feature matching fusion unit introduces low-frequency features and dynamically adjusts the matching weights to provide more high-frequency details, including the following steps:

[0049] right and Divide into local windows and The vector is flattened into a pixel vector within the window, focusing on the feature associations of local regions. Then, the similarity of all positions within the window is calculated using Euclidean distance to obtain a distance matrix, which quantifies the local differences between high and low frequencies.

[0050] For each Window position, filter The most similar position in the window is obtained by using the top-k index, then the corresponding low-frequency features are extracted and reshaped to match the input. With the same spatial dimensions, a local matching association is obtained between the high-frequency features and low-frequency features corresponding to each channel. ;

[0051] The matched features are then concatenated with the original input features, and feature fusion is achieved through low-rank convolution, providing more robust input features for the attention module. .

[0052] In the decoding stage, the inverse wavelet transform (IDWT) reconstructs the multi-band features in the spatial domain. The process is described as follows:

[0053]

[0054] Represents the reconstructed image. This represents the enhanced frequency domain component.

[0055] Compared with the prior art, the beneficial effects of the present invention are:

[0056] 1. This invention solves the problems of traditional methods that rely too much on pixel space processing and insufficient utilization of frequency domain information, thereby enhancing underwater shellfish larvae images.

[0057] 2. Two key modules were invented: one is a low-frequency Mamba processing module for color distortion of low-frequency components, which accurately corrects global color by enhancing long-range dependency modeling capabilities; the other is a high-frequency window enhancement module for detail blurring of high-frequency components, which combines windowed attention and low-rank convolution to achieve adaptive weighting and refined extraction of high-frequency details.

[0058] 3. Experiments show that it exhibits excellent augmentation effects on existing datasets. Attached Figure Description

[0059] Figure 1This is a schematic diagram of the overall network structure of the present invention.

[0060] Figure 2 This is a schematic diagram of a low-frequency Mamba processing module.

[0061] Figure 3 This is a schematic diagram of the high-frequency window enhancement module.

[0062] Figure 4 This is a chart showing the results. Detailed Implementation

[0063] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0064] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention.

[0065] like Figure 1 This is a schematic diagram of the overall network structure of the present invention, illustrating the connection relationship and data flow of the wavelet transform-based U-Net, the low-frequency Mamba processing module, and the high-frequency window enhancement module. The present invention employs a wavelet transform-based U-Net method to perform frequency band separation through a filter bank, thereby achieving complete preservation of multiple frequency bands. Discrete wavelet transform (DWT) achieves multi-scale feature separation through an orthogonal filter bank for a given image. By using Haar wavelet basis functions for decomposition, and through cascading and tensor slicing operations, we can obtain four sets of subband features:

[0066]

[0067] Among them, low-frequency components Encoding global color distribution and illumination characteristics, high-frequency components This captures edge and texture details in the vertical, horizontal, and diagonal directions, respectively. Based on this, a three-layer wavelet decomposition cascaded downsampling architecture is designed, and frequency domain information enhancement is achieved through an interactive dual-branch framework. In the low-frequency enhancement path, the low-frequency Mamba processing module performs cascaded multi-scale long-range feature fusion, modeling cross-level spatial dependencies to fuse image downsampling information. The low-frequency components are enhanced. In the high-frequency optimization path, the fused high-frequency components are refined through a high-frequency window enhancement module under the constraint of low-frequency information. In the decoding stage, inverse wavelet transform (IDWT) is used to reconstruct the multi-band features in the spatial domain. The process can be described as follows:

[0068]

[0069] Represents the reconstructed image. Representing the enhanced frequency domain components, this framework, through frequency band decoupling, dual-path interaction, and full-information reconstruction mechanisms, effectively suppresses feature loss caused by traditional downsampling while inheriting the advantages of U-shaped network topology, achieving high-fidelity restoration of image content. Specific module descriptions will be elaborated below:

[0070] like Figure 2 This is a schematic diagram of the low-frequency Mamba processing module, illustrating its computation process and structure. In this invention, the low-frequency Mamba processing module uses layer normalization to process the input... Stable feature distribution is obtained The Mamba fusion scan block captures long-range structural dependencies between pixels through selective scanning operators, introduces random depth layers to improve training stability while avoiding overfitting, and utilizes 3×3 depthwise convolutions and gating mechanisms in a feedforward network to enhance spatial local interactions of features. Finally, the enhanced low-frequency features are output by a normalization layer. .

[0071] Input features After entering the Mamba fusion scan block, it is projected to the extended dimension via a linear layer and split into two branches. and In the first branch, features After adjusting the dimensions, the feature channels become ,in As an expansion factor, the dimension-adjusted features are subjected to depthwise convolution to extract local features, and then... The activation introduces nonlinear features, which are then combined with a Selected State-Space Model (SSGM) to achieve efficient spatial feature modeling through a dynamic scan path and attention fusion mechanism. Finally, the model is mapped back to the original number of channels through a linear layer. This branching process can be represented as:

[0072]

[0073] in, Indicates the normalization layer. This indicates the selection of a state-space model. Represents a function, This represents depthwise separable convolution. Indicates a linear layer;

[0074] In the second branch, from the projection split ,go through The activation process can be represented as:

[0075]

[0076] The two branches are finally accumulated by Hadamard, and then passed through a linear layer to produce an output of the same size as the input. The overall input-output relationship of the Mamba fused scan block can be represented as:

[0077]

[0078] in, This represents the Hadamard product. This indicates a linear layer.

[0079] The state-space model is chosen to dynamically perceive image degradation differences, comprehensively capture multi-dimensional features, and efficiently fuse key information to achieve precise enhancement of underwater shellfish larvae images. Its key components are the multi-scale feature scanning module and the shuffling attention module. The multi-scale feature scanning module converts 2D image features into 1D sequences and performs multi-scale processing, followed by long-distance dependency modeling. The shuffling attention module enhances key features in underwater shellfish larvae images and suppresses noise and redundant information through inter-channel information interaction and adaptive weight allocation. Specifically:

[0080] The multi-scale feature scanning module dynamically generates four parallel scanning paths based on the resolution of the input image. Each path adopts four different scanning directions: forward, reverse, transposed forward, and transposed reverse. It introduces two core parameters, scanning length and offset length, to achieve selective multi-scale feature extraction, covering information from local details to global structure. Subsequently, based on the scanning path, the 2D feature map is converted into a 1D sequence, and the long-distance dependencies of the sequence features are further captured.

[0081] The shuffling attention module will input features After splicing along the channel dimension, they are all divided into Group, in this example Through dimensional restructuring The channels of the group are mixed together to obtain This shuffling operation allows features from different groups to influence each other, enhancing cross-channel information flow. Next, we perform a non-linear transformation on the compressed features using a global average pooling layer and grouped convolutions to generate channel weights. The generated channel weights are applied to the shuffled features to selectively enhance the features, resulting in selective enhancement. This process can be represented as:

[0082]

[0083]

[0084]

[0085]

[0086] in, Channel splicing operation For channel mixed washing operation, For average pooling layer operations, For convolution operations, For learnable convolution kernel parameters, This is the Sigmoid activation function.

[0087] We regrouped the shredded channels, preserving the original organizational structure of the features while retaining the information exchange between channels.

[0088]

[0089] in, Features after regrouping Regroup the channels.

[0090] like Figure 3 This diagram illustrates a high-frequency window enhancement module, showcasing the process of cross-band feature injection and frequency band reconstruction fusion. Addressing the issue that details such as edges and textures in the high-frequency components of underwater mollusk larvae images are easily affected by scattering and noise interference, this invention proposes a high-frequency window enhancement module based on windowed attention and low-rank convolution. This module achieves accurate capture and efficient enhancement of high-frequency details through a collaborative design of local window partitioning, cross-feature matching, and lightweight convolutional transformation. First, a window partitioning strategy maps high-frequency features to a local window space, reducing the computational complexity of attention. Second, low-frequency features are introduced to enrich high-frequency features, and key details are filtered through feature matching. Furthermore, low-rank matrix factorization is used to optimize the convolution operation, enhancing the ability to capture effective correlations while compressing model parameters.

[0091] The window channel attention mechanism first focuses on high-frequency components. Normalized high-frequency features With enhanced low-frequency characteristics Divide the window into 8x8 sections to obtain local window features. Let ... The query (Q), key (K), and value (V) are generated through 1×1 convolution, and local spatial associations are enhanced through 3x3 depthwise convolution. A cross-feature matching fusion unit is introduced to filter nearest neighbor matching results to optimize the feature representation of Q. Subsequently, the channel attention within the window is calculated through a multi-head self-attention mechanism.

[0092]

[0093] in, For channel attention output, For the attention mechanism, the query matrix, key matrix, and value matrix are... For function, It is a coefficient that controls the size of the Q-K dot product.

[0094] We then obtain the output of the i-th window through residual fusion. and all window features By stitching back the feature map to its original size, we obtain Finally, we unify the channel features through a feedforward network containing 1x1 convolutions to obtain the final output. .

[0095]

[0096]

[0097]

[0098] in, For cross-feature matching and fusion units; It is a feedforward network;

[0099] Cross-feature matching fusion units can provide more high-frequency details by introducing low-frequency features and dynamically adjusting matching weights. We... and Divide into local windows and This is done by flattening the vector into pixel vectors within a window, focusing on feature associations in local regions, and then using Euclidean distance to calculate the similarity of all positions within the window, obtaining a distance matrix that quantifies local differences between high and low frequencies. For each... Window position, filter The most similar position in the window is obtained by using the top-k index, then the corresponding low-frequency features are extracted and reshaped to match the input. With the same spatial dimensions, a local matching association is obtained between the high-frequency features and low-frequency features corresponding to each channel. The matching features are then concatenated with the original input features, and feature fusion is achieved through low-rank convolution, providing more robust input features for the attention module. This process can be represented as:

[0100]

[0101]

[0102] in, It is used to calculate the Euclidean distance between two window positions, representing the quantization of the similarity of positions within the window; The operation is performed to obtain the most similar value; For feature matching operations; For feature splicing; It is a low-rank matrix.

[0103] The overall process of this invention is described as follows: This invention proposes a multi-band UIE network based on the U-Net architecture. The network architecture mainly consists of three parts: a U-shaped network based on wavelet transform, a low-frequency processing module, and a high-frequency enhancement module. Specifically, we first decompose the input features into low-frequency components and three high-frequency components using discrete wavelet transform, and perform global dependency modeling using an improved Mamba structure. The high-frequency branch captures local details through a windowed Transformer. The low-frequency branch achieves long-range dependency modeling through a low-frequency Mamba processing module. The high-frequency branch uses a high-frequency window enhancement module to process the fused high-frequency components, enhancing detail recovery. In the decoding stage, cross-layer feature fusion is achieved through inverse wavelet transform, and channel-aware upsampling is introduced to maintain frequency band consistency. Furthermore, we achieve cross-band feature interaction in the encoding and decoding stages through cross-band feature injection and frequency band reconstruction fusion, thereby further improving our enhancement effect on degraded images of underwater shellfish larvae.

[0104] In the experiment, a self-constructed shellfish larvae dataset was used to test the method of this invention. This dataset contains 2000 pairs of high-quality shellfish larvae images, covering various typical shellfish larval developmental stages. The creation and collection process was systematic and rigorous, as follows: The data collection area selected typical shellfish farming areas and large-scale shellfish seedling greenhouses in the natural intertidal zone along the Yellow and Bohai Seas of my country. This environment is a semi-artificial controlled environment with relatively stable parameters such as water salinity and temperature. The collection time was concentrated in different light intensity periods each day: 6:00-8:00 am, 11:00-13:00 pm, and 16:00-18:00 pm. Several difficulties were encountered during the collection process. First, the lighting conditions were complex and variable, and there were also local uneven lighting problems in the seedling greenhouses, which increased the difficulty of controlling the consistency of image collection. Second, the fluctuation of feed concentration interfered with the data collection. Too high a concentration of feed would obscure the larval morphology, while too low a concentration would lead to a decrease in larval activity. Moreover, the difference in feed types would affect the light transmittance of the water. To ensure the scientific rigor and accuracy of the experiment, a systematic control dataset processing procedure was designed: data screening adopted a dual mechanism of quality screening and diversity screening, removing invalid samples and retaining 1800 high-quality valid samples to ensure a balanced sample ratio and comprehensive scene coverage; the image control set was created to address issues such as uneven lighting and water refraction, and operations such as lighting correction, noise reduction, image alignment and cropping, and water refraction correction were carried out. The sample size was expanded to 2000 images through data augmentation strategies, while the original unprocessed dataset was retained as a baseline dataset. The dataset was then randomly divided into training, validation, and test sets in a 7:2:1 ratio to verify the performance of the method of this invention under different data quality conditions.

[0105] The model was implemented using the PyTorch framework and two NVIDIA GeForce GTX 2080Ti GPUs. The AdamW algorithm was chosen as the optimizer, with β1 and β2 values ​​of 0.9 and 0.99, respectively. The initial learning rate was set to 1×10⁻⁴, and a cosine annealing strategy was used to adjust the learning rate, with a minimum learning rate of 1×10⁻⁷. A total of 300,000 training iterations were performed. Experimental results show that on paired datasets, the proposed method achieves optimal results in both PSNR and SSIM metrics. On unpaired datasets, while not reaching the optimal level, the difference from the best method is small, demonstrating competitiveness. Ablation experiments verified the effectiveness of the high- and low-frequency module combination mode, the combination mode of the low-frequency Mamba processing submodule and the high-frequency window enhancement submodule, and the loss function.

[0106] like Figure 4 As shown in the figure, the larvae are highlighted, demonstrating the significant effectiveness of our method. The enhanced images are more conducive to observation and application.

[0107] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should be considered within the scope of protection of the present invention.

Claims

1. An image enhancement method for underwater sea scallop larvae using a fusion of wavelet transform and Mamba architecture, characterized in that, The following steps are performed to obtain a water fusion wavelet transform and a lower image enhancement network model of the Mamba architecture, which is used to effectively suppress the feature loss caused by traditional downsampling while inheriting the advantages of the U-shaped network topology, realize high-fidelity restoration of image content through frequency band decoupling, double-path interaction and full-information reconstruction mechanism, the method comprising the following steps: A U-Net network based on wavelet transform is established, including: In the encoding stage, a filter bank is constructed based on discrete wavelet transform (DWT) to separate the frequency bands of multi-scale features; a three-layer wavelet decomposition cascade downsampling architecture is constructed, and the information of multi-frequency band features is enhanced through an interactive double-branch framework, which includes a low-frequency Mamba processing module and a high-frequency window enhancement module; wherein the low-frequency Mamba processing module is used to process low-frequency component color distortion and high-frequency component detail blur; the high-frequency window enhancement module is used for cross-frequency feature injection and frequency band reconstruction fusion; In the decoding stage, the inverse wavelet transform (IDWT) is used to reconstruct the multi-band features in the spatial domain, and the target image is obtained Based on the above U-Net network enhanced visible light reconstruction image .

2. The method according to claim 1, wherein the method is characterized by, The discrete wavelet transform (DWT) performs the following steps: For a given target image , using Haar wavelet basis function for decomposition, through cascading operation and tensor slice operation, four groups of sub-band features are obtained: wherein the low frequency component , the encoded global color distribution and illumination characteristics, the high frequency component , respectively capture the edge and texture details in vertical, horizontal and diagonal directions.

3. The method according to claim 1, wherein the method is characterized by, The interactive double-branch framework of the sampling architecture of each layer of wavelet decomposition includes: In the low-frequency enhancement path, a low-frequency Mamba processing module is adopted for cascaded multi-scale long-range feature fusion, and information enhancement is strengthened on the low-frequency components fused with picture down-sampling information by modeling cross-level spatial dependence. ​ In the high-frequency optimization path, the high-frequency window enhancement module is used to realize detail refinement of the fused high-frequency components under the constraint of low-frequency information.

4. The method of claim 3, wherein the method is characterized by, The low-frequency Mamba processing module normalizes the input by layer Stable feature distribution is obtained The Mamba fusion scanning block captures long-distance structure dependencies between pixels by a selective scanning operator, introduces random deep layers, and uses deep convolution and gating mechanisms of the feedforward network to strengthen spatial local interactions of features, and finally outputs enhanced low-frequency features by a normalization layer .

5. The method of claim 4, wherein the method is characterized by, Input features Upon entering the Mamba fusion scan block, the meridians are projected to an expanded dimension via a linear layer and split into two branches and ; In the first branch, the feature After adjusting the dimension, the feature channel becomes where is the expansion factor, the dimension-adjusted feature is extracted for local features through deep convolution, and then The activation introduces nonlinear features, and then combines the state space model (SSGM) to realize efficient spatial feature modeling through dynamic scan path and attention fusion mechanism, and finally maps back to the original channel number through the linear layer; the branch process is represented as: wherein, denotes a normalization layer, denotes a selection state space model, denotes a function, denotes a depthwise separable convolution, denotes a linear layer; In the second branch, the projection split from the first branch is , activated, the process is represented as: ​ The two branches are finally aggregated by Hadamard product and after that, through a linear layer, an output of the same size as the input is produced The overall input-output relationship of the Mamba fusion scan block is represented as wherein denotes the Hadamard product, denotes a linear layer.

6. The method of claim 5, wherein the method is characterized by, The selection state space model includes two modules: a multi-scale feature scanning module and a shuffle attention module; The multi-scale feature scanning module dynamically generates four parallel scanning paths according to the resolution of the input image, each path uses four different scanning directions: forward, reverse, transposed forward, and transposed reverse, and introduces scanning length and offset length parameters to realize selective multi-scale feature extraction, obtaining information covering from local details to global structure; then the 2D feature map is converted into a 1D sequence according to the scanning path, and the long-distance dependence of the sequence features is further captured; The shuffle attention module shuffles the input features After concatenation in the channel dimension, the features are evenly divided into groups, and mixed with each other through dimension reorganization to obtain , enhancing cross-channel information flow. The compressed features are nonlinearly transformed by a global average pooling layer and a grouped convolution to generate channel weights The generated channel weights are applied to the shuffled features to selectively enhance the features to obtain selective enhancement The channel weights are applied to the shuffled features to selectively enhance the features to obtain selective enhancement 7. The method of claim 6, wherein the method is a method of image enhancement of underwater polychaete larvae using a fusion of wavelet transform and Mamba architecture. The processing process of the shuffle attention module is represented as: wherein, channel concatenation operation, channel shuffle operation, average pooling layer operation, convolution operation, learnable convolution kernel parameters, Sigmoid activation function; Re-grouping the shuffled channels: wherein, the re-grouped features, is a channel re-grouping operation.

8. The method of claim 1, wherein the method is a method of enhancing images of sea scallop larvae underwater using fusion wavelet transform and Mamba architecture. The high-frequency window enhancement module first maps the high-frequency features to the local window space through the window partition strategy, reducing the complexity of attention calculation; secondly, it introduces low-frequency features to enrich high-frequency features, and selects key details through feature matching; wherein the low-rank matrix decomposition is used to optimize the convolution operation, which can compress the model parameters while strengthening the effective correlation capture ability; including the following steps: First, high-frequency components normalized high-frequency features and enhanced low-frequency features are divided into 8x8 windows to obtain local window features , i represents the i-th window, focusing on local context and reducing global redundant calculation; Within the window, the features are generated by 1x1 convolution (Q, K, V), and the local spatial correlation is enhanced by 3x3 deep convolution; and a cross-feature matching fusion unit is introduced to filter the nearest neighbor matching results to optimize the feature expression of Q; then, the channel attention within the window is calculated by a multi-head self-attention mechanism: wherein, is the channel attention output, is the query matrix, the key matrix and the value matrix in the attention mechanism, is a function, is a coefficient controlling the size of the Q, K dot product; The output of the i-th window is then obtained by residual fusion , and all window features are stitched back to the original feature map size to obtain , Finally, the channel features are unified by a feed-forward network containing 1x1 convolutions to obtain the final output ; wherein, is a cross-feature matching fusion unit; is a feedforward network.

9. The method of claim 8, wherein the method is characterized by, The cross-feature matching fusion unit dynamically adjusts the matching weight by introducing low-frequency features to provide more high-frequency details, including the following steps: to and are divided into local windows and respectively, flattened into pixel vectors within the window, focusing on the feature correlation of local regions, and then calculating the similarity of all positions within the window using the Euclidean distance to obtain the distance matrix, quantifying the local differences between high and low frequencies; For each window position, filter the most similar positions in the window, get the index by topk, extract the corresponding low-frequency features, and reshape them to the same spatial size as the input , get a local matching correlation of each channel corresponding to the high-frequency features and low-frequency features ; Then the matching features are spliced with the original input features, and the feature fusion is realized through low-rank convolution to provide more robust input features for the attention module .

10. The method of claim 1, wherein the method is an underwater polychaete larvae image enhancement method that fuses wavelet transform and Mamba architecture, characterized by, In the decoding stage, the inverse wavelet transform (IDWT) is used to reconstruct the multi-frequency band features in the spatial domain, and its process is represented as: denotes the reconstructed image, denotes the enhanced frequency domain component.

Images