Underwater sea cucumber image recognition method and recognition system

Abstract

The invention discloses an underwater sea cucumber image recognition method based on a lightweight attention model, which can significantly improve the accuracy of underwater sea cucumber recognition by improving a YOLOv11 model and introducing an LCCA algorithm, and has stronger robustness for small target and background interference especially in a complex underwater environment. Meanwhile, by adopting the depth separable convolution and attention mechanism, the model parameter quantity is reduced, the calculation efficiency is improved, and the method is suitable for real-time operation in the underwater environment with limited resources.

Description

Technical Field

[0001] This invention relates to the field of underwater sea cucumber recognition technology, and in particular to an underwater sea cucumber image recognition method based on a lightweight attention model. Background Technology

[0002] As an important marine economic resource, the monitoring and identification of sea cucumbers in underwater aquaculture and wild environments is of great significance.

[0003] Traditional methods for identifying underwater sea cucumbers mainly rely on manual observation or simple image processing techniques. However, the complex and variable underwater environment, including insufficient light, turbid water, and background interference, severely impacts the accuracy of image acquisition and target identification. Most existing monitoring systems can only provide image acquisition and basic environmental parameter monitoring, lacking the ability to accurately identify underwater sea cucumbers.

[0004] Therefore, it is necessary to provide a new technical solution to solve the above problems. Summary of the Invention

[0005] To address the aforementioned technical issues, this application provides an underwater sea cucumber image recognition method based on a lightweight attention model, which can significantly improve the accuracy of underwater sea cucumber recognition, especially in complex underwater environments, and is more robust to small targets and background interference.

[0006] A method for underwater sea cucumber image recognition includes: Several underwater image data containing sea cucumbers were collected using an underwater camera, covering different lighting conditions and background environments. The quality of underwater sea cucumber images was improved by using local color correction and contrast adaptation methods, resulting in enhanced underwater sea cucumber images and an enhanced sea cucumber image dataset. The enhanced sea cucumber image data was augmented to obtain an augmented dataset; The images in the augmented dataset are labeled with targets and divided into training, validation and test sets according to the proportions. An attention mechanism was embedded as a feature enhancement module into the feature extraction network of the YOLOv11 model to establish a sea cucumber recognition model, and the sea cucumber recognition model was trained. The input image data is enhanced using local color correction and contrast adaptation methods. The enhanced image data is then automatically recognized using a trained sea cucumber recognition model, and the recognition results are visualized.

[0007] Preferably, the method of improving the image quality of underwater sea cucumbers using local color correction and contrast adaptation includes: By applying Gaussian blur and smoothing operations to the image at different levels, a series of image sequences with progressively lower resolution but retaining key image features are generated, thus constructing a pyramid-like structure. Statistically driven color compensation and automatic white balance based on channel differences utilize the differences between the RGB channels of an image to estimate the color of the light source, thereby achieving local color correction of underwater sea cucumber images. An improved contrast fusion strategy is proposed that combines brightness adaptation and local variance constraint. By dynamically adjusting the balance between local image features and global brightness, the distinction between sea cucumbers and background is enhanced.

[0008] Preferably, in the process of generating a series of image sequences with progressively lower resolution but retaining key image features by performing Gaussian blur and smoothing operations on the image at different levels, and constructing a pyramid-shaped structure, the number of layers in the constructed pyramid-shaped structure is 3. The hierarchical expression of the Gaussian pyramid is as follows: ; In the formula, This represents the image of the (k+1)th layer of the Gaussian pyramid. This represents the image of the k-th layer of the Gaussian pyramid. Indicates a downsampling operation; The standard deviation is expressed as Gaussian kernel; The larger the value, the stronger the blurring effect and the more image details are lost. These are commonly used parameters, designed to strike a balance between blurring and preserving key information; In generating a series of images with progressively decreasing resolution but retaining key image features, high-frequency details are preserved and complex frequency band recombination is avoided by using the high-frequency residuals of the images. The formula for calculating the high-frequency residuals of the images is as follows: ; In the formula, The high-frequency residual of the image represents the difference between the pixel values ​​of the original image and the image after Gaussian filtering. Represents the original image. This represents the image after Gaussian filtering the original image.

[0009] Preferably, the statistically driven color compensation, based on automatic white balance with channel differences, estimates the color of the light source using the differences between the RGB channels of the image, including: Calculate the mean value for each color channel: ; In the formula, R represents the red channel, G represents the green channel, and B represents the blue channel. The average value for each color channel is denoted as , where It can be the red channel, green channel, or blue channel; N is the total number of pixels in the image; To calculate the sum of all pixel values ​​in this channel; Indicates the coordinates in the image are Pixels in channel Pixel values ​​on; Calculate the adaptive compensation coefficient: ; In the formula, Indicates the weighting coefficient of the color channel; These are the average values ​​for the red, green, and blue channels, reflecting the overall brightness level of each channel; The average brightness of the brightest channel in the image; It is a very small constant used to avoid the denominator being 0; Perform color correction output: ; ; In the formula, Indicates the coordinates in the image are Pixels in channel Pixel values ​​on; Indicates the weighting coefficient of the color channel; This represents the weighted pixel value; This represents the set of pixel values ​​for the three channels after weighted adjustment. It is a clipping function, because The value may exceed the range of 0-255. Use this function to limit the pixel value to between 0-255.

[0010] Preferably, the improvement of the contrast fusion strategy by combining brightness adaptation and local variance constraint, which enhances the distinguishability of sea cucumbers and background by dynamically adjusting the balance between local image features and global brightness, includes: The values ​​of the three RGB channels in the image are converted into a single brightness value by weighted summation. ; In the formula, This is the brightness value. The value of the red channel. The value for the green channel. The value of the blue channel; Dynamically adjusts local contrast and white balance of an image, dynamically adjusting white balance gain based on local image features to achieve more accurate color reproduction: ; In the formula, The output image after fusion; The weights that influence the importance of device input characteristics; This indicates the relative importance of low-frequency information in the original image; This represents pixel values ​​that contain high-frequency detail information and have undergone color correction. Represents low-frequency information such as basic color balance and brightness distribution in an image, as well as an attention map; in, ; ; In the formula, To determine the weights that influence the importance of device input characteristics, A value close to 1 indicates that high-frequency information should dominate in subsequent processing; and Complementarity essentially reflects the relative importance of low-frequency information in the original image; The variance of the original image parameters. The variance of the high-frequency residuals of the image; Used to measure the proportion of high-frequency information in the total change of the original image; The fusion results are subjected to limited contrast histogram equalization: ; In the formula, For the final output enhanced image; To output the image, A contrast limit threshold is used to control the cumulative proportion of pixel values ​​in the histogram, avoiding areas that are too bright or too dark. To divide the image into For each sub-block, histogram equalization is performed separately.

[0011] Preferably, the attention mechanism is divided into a multi-scale frequency domain decomposition stage, a physical guided modeling stage, and a dynamic degradation perception fusion stage. In the multi-scale frequency domain decomposition stage, the image is accurately decomposed into multi-scale and multi-directional sub-bands through hybrid wavelet packet transform, thereby accurately extracting different frequency band features containing low-frequency color, mid-frequency edge and high-frequency texture. In the physical-guided modeling stage, the underwater light attenuation model is combined with the Rayleigh scattering model to quantify the attenuation degree of light signals and the intensity of noise interference in different frequency bands, thereby generating corresponding physical prior weights. In the dynamic degradation perception fusion stage, a dual-path scoring network is designed to evaluate the semantic saliency and degradation degree of sub-bands respectively. Based on the degradation degree of local regions, the attention weight of each frequency band is adaptively adjusted to achieve a balance between color correction and detail enhancement.

[0012] Preferably, the hybrid wavelet packet transform is as follows: a directional filter bank is introduced based on traditional wavelet decomposition to enhance the capture of diagonal details. The decomposition formula is as follows: ; in, For the input image, This is the first low-frequency sub-band. For the first layer of horizontal high-frequency sub-band, This is the first layer of vertical high-frequency sub-band. This is the first layer of diagonal high-frequency sub-bands; , , , It is a decomposition operator in the hybrid wavelet packet transform that targets different directional and frequency characteristics, used to decompose the input image. Extract the corresponding sub-band features.

[0013] The preferred calculation formula for the underwater light attenuation model is as follows: ; In the formula, Indicates the decay weight. The scale level representing the decomposition. Represents color channels; The attenuation coefficient is wavelength-dependent. Indicates the first The squared gradient of the low-frequency sub-band of the layer is used to reflect the degree of drastic change in the brightness of the region; The formula for calculating Rayleigh scattering is: ; In the formula, Indicates noise weight, The scale level representing the decomposition. Represents color channels. Corresponding to the wavelength of light in the color channel It represents the L1 norm of the diagonal high-frequency subband, reflecting the intensity of high-frequency details at that location.

[0014] Preferably, a dual-path scoring network is designed to evaluate the semantic saliency and degradation degree of subbands, including: Creating semantically saliency paths: Extracting sub-band features using lightweight convolutional layers and outputting saliency scores: ; In the formula, For semantic saliency score, For the first Low-frequency sub-band For the first Layered horizontal high-frequency sub-band, For the first Vertical high-frequency sub-bands For the first Diagonal high-frequency subbands This is a 3×3 convolution operation used to fuse multi-subband features and extract local correlation information; The activation function maps the output value to the range of 0-1, resulting in a normalized significance score. Create a degradation path and fuse attenuation and noise information using physical weights: ; In the formula, Score the degree of degradation; The hyperbolic tangent activation function maps the output value to the range of -1 to 1, representing the degree of normalization degradation. Indicates the decay weight. Indicates noise weight; The final attention map is generated by weighting and fusing semantic saliency with the degree of degradation. ; In the formula, This is an attention weight map; These are adaptive parameters that are dynamically adjusted based on the global image degradation level. , The grayscale entropy of the input image; The normalization function ensures that the sum of the weights of all pixels in the weight map is 1, highlighting the relative proportion of important regions. Multi-scale feature reconstruction is performed, and attention weights are applied to each sub-band before the inverse hybrid wavelet packet transform: ; In the formula, The features are those obtained after weighted fusion; This is a hybrid wavelet packet inverse transform used to reconstruct high-frequency subbands into detailed feature maps; This is the Hadamard product, used to weight the attention weights on the features; Enhanced image output through cross-scale fusion: ; In the formula, For the final enhanced image; This indicates that the upsampled features at three scales are summed and overlaid. This indicates a skip connection, directly importing the original image. This information helps prevent the loss of underlying details during multi-scale fusion.

[0015] According to another aspect of the present invention, an underwater sea cucumber image recognition system is provided for performing the aforementioned underwater sea cucumber image recognition method, comprising: The underwater image acquisition module uses an underwater high-definition camera to acquire images of the underwater environment in real time. The image preprocessing module uses local color correction and contrast adaptation to enhance the acquired image. The recognition module includes a trained YOLOv11 model with an attention mechanism, which extracts features from sea cucumber images and performs recognition by pre-training the model. The results output module uses a visual interface designed with PyQT5 technology to display the recognition results.

[0016] Compared with the prior art, this application has at least the following beneficial effects: 1. This invention significantly improves the accuracy of underwater sea cucumber identification by improving the YOLOv11 model and introducing the LCCA algorithm, especially in complex underwater environments, where it is more robust to small targets and background interference.

[0017] 2. This invention employs depthwise separable convolution and attention mechanisms, reducing the number of model parameters and improving computational efficiency, making it suitable for real-time operation in resource-constrained underwater environments. The image preprocessing module effectively improves underwater image quality, further enhancing recognition performance.

[0018] 3. This invention enhances the separation capability of diagonal details by combining directional filter banks and wavelet packets, avoiding the directional limitations of traditional wavelet transforms. Physical prior weights are generated through light attenuation and scattering models, combined with data-driven saliency scoring to enhance the interpretability and generalization of the model. Furthermore, attention weights are adaptively adjusted based on image entropy and local gradients, prioritizing noise suppression in severely degraded regions and enhancing details in mildly degraded regions. Attached Figure Description

[0019] The following sections will describe some specific embodiments of the invention in detail by way of example and not limitation, with reference to the accompanying drawings. The same reference numerals in the drawings denote the same or similar parts or portions. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. In the drawings: Figure 1 This is a schematic diagram of the overall process of the present invention; Figure 2 The original image that was captured; Figure 3 The image is after LCCA processing; Figure 4 Image of the recognition result. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] like Figure 1 As shown, an underwater sea cucumber image recognition method includes the following steps: S1. Collect several underwater image data containing sea cucumbers and covering different lighting conditions and background environments using an underwater camera.

[0022] Among them, underwater images containing sea cucumbers and covering different lighting conditions and background environments were collected, such as... Figure 2 As shown.

[0023] S2. The quality of underwater sea cucumber images is improved by using the Local Color Correction and Contrast Adaptation (LCCA) method, resulting in enhanced underwater sea cucumber images and an enhanced sea cucumber image dataset.

[0024] Specifically, including: S21. By performing Gaussian blur and smoothing operations on the image at different levels, a series of image sequences with progressively reduced resolution but retaining key image features are generated, thus constructing a pyramid-shaped structure.

[0025] This improvement involves using a fast multi-scale decomposition method and replacing complex filtering with Gaussian pyramids to accelerate detail extraction. Gaussian pyramid filtering is an image multi-scale analysis method based on Gaussian functions. By applying Gaussian blurring and smoothing operations to different levels of the image, it generates a series of images with progressively decreasing resolution, constructing a pyramid-like structure. In each level, the image size is reduced by half while preserving key image features.

[0026] First, a pyramid is constructed by downsampling to generate a 3-layer Gaussian pyramid. The layer expression for the Gaussian pyramid is as follows: ; In the formula, This represents the image of the (k+1)th layer of the Gaussian pyramid. This represents the image of the k-th layer of the Gaussian pyramid. Indicates a downsampling operation; The standard deviation is expressed as Gaussian kernel; The larger the value, the stronger the blurring effect and the more image details are lost. These are commonly used parameters, designed to strike a balance between blurring and preserving key information; In generating a series of images with progressively decreasing resolution but retaining key image features, high-frequency details are preserved and complex frequency band recombination is avoided by using the high-frequency residuals of the images. The formula for calculating the high-frequency residuals of the images is as follows: ; In the formula, The high-frequency residual of the image represents the difference between the pixel values ​​of the original image and the image after Gaussian filtering. Represents the original image. This represents the image after Gaussian filtering the original image.

[0027] S22. Statistically driven color compensation, automatic white balance based on channel differences, uses the differences between the RGB channels of the image to estimate the color of the light source, and realizes local color correction of underwater sea cucumber images.

[0028] The core of this step is to find white or gray spots in the image and achieve white balance by adjusting the channel gain.

[0029] Calculate the mean value for each color channel: ; In the formula, R represents the red channel, G represents the green channel, and B represents the blue channel. The average value for each color channel is denoted as , where It can be the red channel, green channel, or blue channel; N is the total number of pixels in the image; To calculate the sum of all pixel values ​​in this channel; Indicates the coordinates in the image are Pixels in channel Pixel values ​​on; Calculate the adaptive compensation coefficient: ; In the formula, Indicates the weighting coefficient of the color channel; These are the average values ​​for the red, green, and blue channels, reflecting the overall brightness level of each channel; The average brightness of the brightest channel in the image; It is a very small constant used to avoid the denominator being 0; Perform color correction output: ; ; In the formula, Indicates the coordinates in the image are Pixels in channel Pixel values ​​on; Indicates the weighting coefficient of the color channel; This represents the weighted pixel value; This represents the set of pixel values ​​for the three channels after weighted adjustment. It is a clipping function, because The value may exceed the range of 0-255. Use this function to limit the pixel value to between 0-255.

[0030] S23. An improved contrast fusion strategy is implemented by combining brightness adaptation and local variance constraints. By dynamically adjusting the balance between local image features and global brightness, the distinction between sea cucumbers and background is enhanced.

[0031] The values ​​of the three RGB channels in the image are converted into a single brightness value by weighted summation. ; In the formula, This is the brightness value. The value of the red channel. The value for the green channel. This is the value of the blue channel.

[0032] The weighting coefficients are chosen based on the human eye’s sensitivity to different colors. In the formula, the green channel dominates the brightness calculation, while the blue channel contributes relatively little. This coefficient is consistent with the actual situation of underwater images.

[0033] ; ; In the formula, To determine the weights that influence the importance of device input characteristics, A value close to 1 indicates that high-frequency information should dominate in subsequent processing; and Complementarity essentially reflects the relative importance of low-frequency information in the original image; The variance of the original image parameters. The variance of the high-frequency residuals of the image; Used to measure the proportion of high-frequency information in the total change of the original image; Based on the image information obtained above, three key types of information are identified: high-frequency residuals of the image. Pixel values ​​that represent high-frequency detail information and have undergone color correction. This represents low-frequency information such as the image's basic color balance and brightness distribution, as well as the attention map. Attention maps are weighted maps that highlight key regions in an image and represent important details. A value of 0.1 indicates that the attention map accounts for 10% of the fusion process, enabling collaborative integration.

[0034] Dynamically adjusts local contrast and white balance of an image, dynamically adjusting white balance gain based on local image features to achieve more accurate color reproduction: ; In the formula, The output image after fusion; The weights that influence the importance of device input characteristics; This indicates the relative importance of low-frequency information in the original image; This represents pixel values ​​that contain high-frequency detail information and have undergone color correction. It represents low-frequency information such as basic color balance and brightness distribution of an image, as well as attention maps.

[0035] From High-frequency enhancement makes details clearer, from The correction process makes the colors more balanced; from Attention guidance makes the key points more prominent, ultimately resulting in the fused output image. .

[0036] The fusion results are subjected to limited contrast histogram equalization: ; In the formula, For the final output enhanced image; To output the image, A contrast limit threshold is used to control the cumulative proportion of pixel values ​​in the histogram, avoiding areas that are too bright or too dark. To divide the image into The sub-blocks are processed individually, and histogram equalization is performed on each sub-block. The more sub-blocks there are, the more refined the local contrast adjustment can be.

[0037] S3. The enhanced sea cucumber image data is augmented to obtain an augmented dataset.

[0038] S4. Label the images in the augmented dataset and divide them into training, validation and test sets according to the proportions.

[0039] The dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for model training, the validation set is used for model tuning and parameter selection, and the test set is used to evaluate the model's performance and generalization ability.

[0040] S5. The attention mechanism is embedded as a feature enhancement module into the feature extraction network of the YOLOv11 model to establish a sea cucumber recognition model, and the sea cucumber recognition model is trained.

[0041] In the YOLOv11 model, an attention mechanism is introduced and improved. The attention mechanism is a technique in deep learning inspired by how humans allocate attention, aiming to allow the model to dynamically focus on the most critical parts of the input data. It primarily borrows from a human characteristic in information processing: given limited resources, prioritizing attention on the most important information to improve efficiency and accuracy. By dynamically allocating weights, the attention mechanism allows the model to focus on the most important parts of the input data, thereby improving processing efficiency and accuracy.

[0042] To address numerous issues in underwater images, such as color shift, low-frequency blur, high-frequency noise, and dynamic illumination changes, this study proposes a three-stage attention framework: multi-scale frequency domain decomposition, physical-guided modeling, and dynamic degradation perception.

[0043] Specifically, the following steps are included: S51. In the multi-scale frequency domain decomposition stage, the image is accurately decomposed into multi-scale and multi-directional sub-bands through hybrid wavelet packet transform, thereby accurately extracting different frequency band features containing low-frequency color, mid-frequency edge and high-frequency texture.

[0044] First, wavelet transform is performed on the image to obtain low-frequency subbands and horizontal, vertical, and diagonal high-frequency subbands. Then, a directional filter bank is introduced to further decompose the high-frequency subbands, enhancing the ability to capture details in diagonal and other directions. Finally, through multiple rounds of iteration, subband sets of different scales and directions are generated to achieve feature extraction from global to local and multi-dimensional dimensions, thereby accurately extracting different frequency band features including low-frequency color, mid-frequency edges, and high-frequency texture.

[0045] Hybrid wavelet packet transform is achieved by introducing a directional filter bank on the basis of traditional wavelet decomposition to enhance the ability to capture diagonal details. This is because traditional wavelet decomposition mainly filters in the horizontal and vertical directions, which has a strong ability to capture high-frequency information in the horizontal and vertical directions, but a weak ability to capture details in the diagonal direction, and is prone to losing information. Therefore, by designing a filter kernel in the diagonal direction, the capture of information in the 45° and 135° diagonal directions can be enhanced.

[0046] The image decomposition formula is as follows:

[0047] ; in, For the input image, This is the first low-frequency sub-band. For the first layer of horizontal high-frequency sub-band, This is the first layer of vertical high-frequency sub-band. This is the first layer of diagonal high-frequency sub-bands; , , , It is a decomposition operator in the hybrid wavelet packet transform that targets different directional and frequency characteristics, used to decompose the input image. Extract the corresponding sub-band features; where The operator is designed using a second-order difference operator to create a filter kernel that is sensitive to the diagonal direction. By setting a weight of 1 on the diagonal and a weight of 0 on the off-diagonal, when the filter kernel is convolved with the image, the edge pixels in the diagonal direction will produce a high response, thereby highlighting the diagonal details.

[0048] S52. In the physical-guided modeling stage, the underwater light attenuation model is combined with the Rayleigh scattering model to quantify the attenuation degree of light signals and the intensity of noise interference in different frequency bands, thereby generating the corresponding physical prior weights.

[0049] According to Beer-Lambert's law for underwater light propagation, the wavelength-dependent attenuation coefficient is defined as follows: ; In the formula, Corresponding to RGB channels; For scene depth, The initial attenuation coefficient, The absorption coefficient of the water body.

[0050] Local depth is obtained through a pre-trained depth estimation network, and the attenuation weights of each sub-band are calculated to form an underwater light attenuation model: ; ; In the formula, Indicates the decay weight. The scale level representing the decomposition. Represents color channels; The attenuation coefficient is wavelength-dependent. Indicates the first The squared gradient of the low-frequency subband reflects the degree of brightness change in the region; the larger the value, the more obvious the details. This formula can be used to assign attenuation weights to features in different regions and channels, enabling the model to weaken features severely affected by light attenuation in subsequent processing, while retaining effective information of key targets such as sea cucumbers.

[0051] Establish a scattering noise model: Utilizing the inverse relationship between noise intensity and wavelength caused by Rayleigh scattering, define the noise weights for high-frequency subbands, and quantify the high-frequency noise intensity through norm normalization. The calculation formula for the Rayleigh scattering model is as follows: ; In the formula, Indicates noise weight, The scale level representing the decomposition. Represents color channels. Corresponding to the wavelength of light in the color channel This represents the L1 norm of the diagonal high-frequency subband, reflecting the intensity of high-frequency details at that location. This formula is used to reduce the attention given to areas with high noise weights, preventing noise from interfering with the identification of sea cucumber features.

[0052] S53. In the dynamic degradation perception fusion stage, a dual-path scoring network is designed to evaluate the semantic saliency and degradation degree of sub-bands respectively. Based on the degradation degree of local regions, the attention weight of each frequency band is adaptively adjusted to achieve a balance between color correction and detail enhancement.

[0053] Specifically, a dual-path scoring network is designed to evaluate the semantic saliency and degradation degree of subbands, based on the degradation degree of local regions, including: Creating semantically saliency paths: Extracting sub-band features using lightweight convolutional layers and outputting saliency scores: ; In the formula, For semantic saliency score, For the first Low-frequency sub-band For the first Layered horizontal high-frequency sub-band, For the first Vertical high-frequency sub-bands For the first Diagonal high-frequency subbands This is a 3×3 convolution operation used to fuse multi-subband features and extract local correlation information; The activation function maps the output value to the range of 0-1, resulting in a normalized significance score. Create a degradation path and fuse attenuation and noise information using physical weights: ; In the formula, Score the degree of degradation; The hyperbolic tangent activation function maps the output value to the range of -1 to 1, representing the degree of normalization degradation. Indicates the decay weight. Indicates noise weight; The final attention map is generated by weighting and fusing semantic saliency with the degree of degradation. ; In the formula, This is an attention weight map; These are adaptive parameters that are dynamically adjusted based on the global image degradation level. , The grayscale entropy of the input image; The normalization function ensures that the sum of the weights of all pixels in the weight map is 1, highlighting the relative proportion of important regions. Multi-scale feature reconstruction is performed, and attention weights are applied to each sub-band before the inverse hybrid wavelet packet transform: ; In the formula, The features are those obtained after weighted fusion; This is a hybrid wavelet packet inverse transform used to reconstruct high-frequency subbands into detailed feature maps; This is the Hadamard product, used to weight the attention weights on the features; Enhanced image output through cross-scale fusion: ; In the formula, For the final enhanced image; This indicates that the upsampled features at three scales are summed and overlaid. This indicates a skip connection, directly importing the original image. This information helps prevent the loss of underlying details during multi-scale fusion.

[0054] S6. The input image data is enhanced using local color correction and contrast adaptation methods, and the enhanced image data is automatically recognized using the trained sea cucumber recognition model. The recognition results are then visualized.

[0055] The input image data can be in the form of three types: images, videos, and camera video streams. The recognition results are as follows: Figure 4 As shown.

[0056] In the evaluation of the underwater sea cucumber image recognition method of this invention, it was found that mAP@0.5 is a commonly used accuracy evaluation index in target detection tasks. It represents the average accuracy when the intersection-union ratio (IU) threshold is 0.5, and can more comprehensively reflect the accuracy of the model. The larger the mAP value, the better the performance. GFLOPs represents billions of floating-point operations per second. It is an important indicator for measuring system performance and is used to evaluate the computational complexity of deep learning models.

[0057] Table 1 Model Evaluation Table

[0058] Table 1 shows that the system has higher recognition accuracy while greatly reducing computational complexity.

[0059] Based on the same inventive concept, an underwater sea cucumber image recognition system is also provided for performing the above-mentioned underwater sea cucumber image recognition method, comprising: The underwater image acquisition module uses an underwater high-definition camera to acquire images of the underwater environment in real time. The image preprocessing module uses local color correction and contrast adaptation to enhance the acquired image. The recognition module includes a trained YOLOv11 model with an attention mechanism, which extracts features from sea cucumber images and performs recognition by pre-training the model. The results output module uses a visual interface designed with PyQT5 technology to display the recognition results.

[0060] The underwater sea cucumber image recognition method and system of the present invention can be widely used in sea cucumber farming management, marine biological surveys and other fields, and has important economic and social value.

[0061] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0062] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in sequences other than those illustrated or described herein.

[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for underwater sea cucumber image recognition, characterized in that, include: Several underwater image data containing sea cucumbers were collected using an underwater camera, covering different lighting conditions and background environments. The quality of underwater sea cucumber images was improved by using local color correction and contrast adaptation methods, resulting in enhanced underwater sea cucumber images and an enhanced sea cucumber image dataset. The enhanced sea cucumber image data was augmented to obtain an augmented dataset; The images in the augmented dataset are labeled with targets and divided into training, validation and test sets according to the proportions. An attention mechanism was embedded as a feature enhancement module into the feature extraction network of the YOLOv11 model to establish a sea cucumber recognition model, and the sea cucumber recognition model was trained. The input image data is enhanced using local color correction and contrast adaptation methods. The enhanced image data is then automatically recognized using a trained sea cucumber recognition model, and the recognition results are visualized.

2. The underwater sea cucumber image recognition method as described in claim 1, characterized in that, The method for improving underwater sea cucumber image quality using local color correction and contrast adaptation includes: By applying Gaussian blur and smoothing operations to the image at different levels, a series of image sequences with progressively lower resolution but retaining key image features are generated, thus constructing a pyramid-like structure. Statistically driven color compensation and automatic white balance based on channel differences utilize the differences between the RGB channels of an image to estimate the color of the light source, thereby achieving local color correction of underwater sea cucumber images. An improved contrast fusion strategy is proposed that combines brightness adaptation and local variance constraint. By dynamically adjusting the balance between local image features and global brightness, the distinction between sea cucumbers and background is enhanced.

3. The underwater sea cucumber image recognition method as described in claim 2, characterized in that, By performing Gaussian blurring and smoothing operations on images at different levels, a series of image sequences with progressively lower resolution but retaining key image features are generated, and a pyramid-shaped structure is constructed. The pyramid-shaped structure has 3 layers. The hierarchical expression of the Gaussian pyramid is as follows: ； In the formula, This represents the image of the (k+1)th layer of the Gaussian pyramid. This represents the image of the k-th layer of the Gaussian pyramid. Indicates a downsampling operation; The standard deviation is expressed as Gaussian kernel; The larger the value, the stronger the blurring effect and the more image details are lost. These are commonly used parameters, designed to strike a balance between blurring and preserving key information; In generating a series of images with progressively decreasing resolution but retaining key image features, high-frequency details are preserved and complex frequency band recombination is avoided by using the high-frequency residuals of the images. The formula for calculating the high-frequency residuals of the images is as follows: ； In the formula, The high-frequency residual of the image represents the difference between the pixel values ​​of the original image and the image after Gaussian filtering. Represents the original image. This represents the image after Gaussian filtering the original image.

4. The underwater sea cucumber image recognition method as described in claim 3, characterized in that, The statistically driven color compensation, based on automatic white balance using channel differences, estimates the color of the light source by utilizing the differences between the RGB channels of the image, including: Calculate the mean value for each color channel: ； In the formula, R represents the red channel, G represents the green channel, and B represents the blue channel. The average value for each color channel is denoted as , where It can be the red channel, green channel, or blue channel; N is the total number of pixels in the image; To calculate the sum of all pixel values ​​in this channel; Indicates the coordinates in the image are Pixels in channel Pixel values ​​on; Calculate the adaptive compensation coefficient: ； In the formula, Indicates the weighting coefficient of the color channel; These are the average values ​​for the red, green, and blue channels, reflecting the overall brightness level of each channel; The average brightness of the brightest channel in the image; It is a very small constant used to avoid the denominator being 0; Perform color correction output: ； ； In the formula, Indicates the coordinates in the image are Pixels in channel Pixel values ​​on; Indicates the weighting coefficient of the color channel; This represents the weighted pixel value; This represents the set of pixel values ​​for the three channels after weighted adjustment. It is a clipping function.

5. The underwater sea cucumber image recognition method as described in claim 4, characterized in that, The improved contrast fusion strategy, which combines brightness adaptation and local variance constraints to enhance the distinction between sea cucumbers and backgrounds by dynamically adjusting the balance between local image features and global brightness, includes: The values ​​of the three RGB channels in the image are converted into a single brightness value by weighted summation. ； In the formula, This is the brightness value. The value of the red channel. The value for the green channel. The value of the blue channel; Dynamically adjusts local contrast and white balance of an image, dynamically adjusting white balance gain based on local image features to achieve more accurate color reproduction: ； In the formula, The output image after fusion; The weights that influence the importance of device input characteristics; This indicates the relative importance of low-frequency information in the original image; This represents pixel values ​​that contain high-frequency detail information and have undergone color correction. Represents low-frequency information such as basic color balance and brightness distribution in an image, as well as an attention map; in, ； ； In the formula, To determine the weight of the influence of device input characteristics on their importance. A value close to 1 indicates that high-frequency information should dominate in subsequent processing; and Complementarity essentially reflects the relative importance of low-frequency information in the original image; The variance of the original image parameters. The variance of the high-frequency residuals of the image; Used to measure the proportion of high-frequency information in the total change of the original image; The fusion results are subjected to limited contrast histogram equalization: ； In the formula, For the final output enhanced image; To output the image, A contrast limit threshold is used to control the cumulative proportion of pixel values ​​in the histogram, avoiding areas that are too bright or too dark. To divide the image into For each sub-block, histogram equalization is performed separately.

6. The underwater sea cucumber image recognition method as described in claim 1, characterized in that, The attention mechanism is divided into a multi-scale frequency domain decomposition stage, a physical guided modeling stage, and a dynamic degradation perception fusion stage. In the multi-scale frequency domain decomposition stage, the image is accurately decomposed into multi-scale and multi-directional sub-bands through hybrid wavelet packet transform, thereby accurately extracting different frequency band features containing low-frequency color, mid-frequency edge and high-frequency texture. In the physical-guided modeling stage, the underwater light attenuation model is combined with the Rayleigh scattering model to quantify the attenuation degree of light signals and the intensity of noise interference in different frequency bands, thereby generating corresponding physical prior weights. In the dynamic degradation perception fusion stage, a dual-path scoring network is designed to evaluate the semantic saliency and degradation degree of sub-bands respectively. Based on the degradation degree of local regions, the attention weight of each frequency band is adaptively adjusted to achieve a balance between color correction and detail enhancement.

7. The underwater sea cucumber image recognition method as described in claim 6, characterized in that, The hybrid wavelet packet transform is achieved by introducing a directional filter bank on the basis of traditional wavelet decomposition to enhance the ability to capture diagonal details. The decomposition formula is as follows: ； in, For the input image, This is the first low-frequency sub-band. For the first layer of horizontal high-frequency sub-band, This is the first layer of vertical high-frequency sub-band. This is the first layer of diagonal high-frequency sub-bands; , , , It is a decomposition operator in the hybrid wavelet packet transform that targets different directional and frequency characteristics, used to decompose the input image. Extract the corresponding sub-band features.

8. The underwater sea cucumber image recognition method as described in claim 7, characterized in that, The calculation formula for the underwater light attenuation model is as follows: ； In the formula, Indicates the decay weight. The scale level representing the decomposition. Represents color channels; The attenuation coefficient is wavelength-dependent. Indicates the first The squared gradient of the low-frequency sub-band of the layer is used to reflect the degree of drastic change in the brightness of the region; The formula for calculating Rayleigh scattering is: ； In the formula, Indicates noise weight, The scale level representing the decomposition. Represents color channels. Corresponding color channel wavelengths It represents the L1 norm of the diagonal high-frequency subband, reflecting the intensity of high-frequency details at that location.

9. The underwater sea cucumber image recognition method as described in claim 8, characterized in that, Design a dual-path scoring network to evaluate the semantic saliency and degradation degree of subbands, including: Creating semantically saliency paths: Extracting sub-band features using lightweight convolutional layers and outputting saliency scores: ； In the formula, For semantic saliency score, For the first Low-frequency sub-band For the first Layered horizontal high-frequency sub-band, For the first Vertical high-frequency sub-bands For the first Diagonal high-frequency subbands This is a 3×3 convolution operation used to fuse multi-subband features and extract local correlation information; The activation function maps the output value to the range of 0-1, resulting in a normalized significance score. Create a degradation path and fuse attenuation and noise information using physical weights: ； In the formula, Score the degree of degradation; The hyperbolic tangent activation function maps the output value to the range of -1 to 1, representing the degree of normalization degradation. Indicates the decay weight. Indicates noise weight; The final attention map is generated by weighting and fusing semantic saliency with the degree of degradation. ； In the formula, This is an attention weight map; These are adaptive parameters that are dynamically adjusted based on the global image degradation level. , The grayscale entropy of the input image; The normalization function ensures that the sum of the weights of all pixels in the weight map is 1, highlighting the relative proportion of important regions. Multi-scale feature reconstruction is performed, and attention weights are applied to each sub-band before the inverse hybrid wavelet packet transform: ； In the formula, The features are those obtained after weighted fusion; This is a hybrid wavelet packet inverse transform used to reconstruct high-frequency subbands into detailed feature maps; This is the Hadamard product, used to weight the attention weights on the features; Enhanced image output through cross-scale fusion: ； In the formula, For the final enhanced image; This indicates that the upsampled features at three scales are summed and overlaid. This indicates a skip connection, directly importing the original image. This information helps prevent the loss of underlying details during multi-scale fusion.

10. An underwater sea cucumber image recognition system, characterized in that, The method for performing underwater sea cucumber image recognition according to any one of claims 1-9 includes: The underwater image acquisition module uses an underwater high-definition camera to acquire images of the underwater environment in real time. The image preprocessing module uses local color correction and contrast adaptation to enhance the acquired image. The recognition module includes a trained YOLOv11 model with an attention mechanism, which extracts features from sea cucumber images and performs recognition by pre-training the model. The results output module uses a visual interface designed with PyQT5 technology to display the recognition results.

Images