Method and system for acquiring litchi meat ice crystal structure image under low-frequency electric field

Abstract

The invention provides a method and system for acquiring an ice crystal structure image of litchi meat in a low-frequency electric field, and the method comprises the steps: carrying out the magnetic resonance scanning of the litchi which is processed by the low-frequency electric field and is stored under an ice temperature condition, and obtaining the k-space data of the litchi meat; performing Fourier transform on the k-space data to obtain image space data; inputting the amplitude information and the phase information of the image space data into a pre-trained convolutional neural network model as two channels to obtain a litchi meat ice crystal structure image with enhanced ice crystal structure definition; wherein the model comprises a convolution attention module which is used for enhancing the sensing ability of the litchi ice crystal area. In the model training process, a composite loss function is adopted, the weight of each item in the composite loss function is dynamically optimized and adjusted by using a particle swarm optimization algorithm, k-space data in litchis are acquired in combination with a magnetic resonance imaging technology, image space data are processed and reconstructed by using a convolutional neural network model, and the image space data are acquired. And generating a high-definition ice crystal structure image.

Description

Technical Field

[0001] The present invention relates to the technical field of food preservation technology, and in particular to a method and system for acquiring an image of ice crystal structure of lychee flesh under a low-frequency electric field. Background Art

[0002] Lychees are renowned for their unique flavor and rich nutritional value. Lychee fruits and their tissues are rich in a variety of physiologically active substances, including polyphenols, flavonoids, and polysaccharides. Lychees are typical climacteric fruits, ripening rapidly after harvest and exhibiting a vigorous metabolic activity. Their unique pericarp structure makes them susceptible to water loss and browning. These characteristics result in a very short shelf life for lychees, making them susceptible to rot, deterioration, flavor degradation, and nutrient loss during harvesting, transportation, storage, and marketing. This severely impacts their commercial value and limits the further development of the lychee industry.

[0003] Low-temperature storage is a common and effective method to inhibit the physiological metabolism of fruits and vegetables after harvest, delay aging, inhibit microbial activity, and thus extend the shelf life. Depending on the temperature range, low-temperature preservation can be divided into refrigeration (usually above 0°C), ice temperature (between 0°C and tissue freezing point), micro-freezing, and freezing. Among them, ice temperature preservation technology has unique advantages because it can maintain food in a state close to the freezing point but not frozen. Compared with traditional refrigeration, it can significantly inhibit enzyme activity and microbial growth at lower temperatures, while avoiding ice crystal damage and quality deterioration caused by freezing.

[0004] Although ice temperature preservation has shown potential in other foods, research on ice temperature preservation of litchi is currently insufficient. In particular, under ice temperature conditions, the internal tissue will inevitably face the problem of ice crystal formation. The size, shape, distribution of ice crystals and their interaction with cell structure are directly related to the integrity of the fruit's tissue structure, juice retention capacity and ultimate edible quality. However, a deep understanding of the microscopic mechanism of ice crystal formation in litchi pulp under ice temperature conditions is crucial for optimizing the preservation process and maximizing the quality of litchi. Existing conventional detection methods often find it difficult to non-destructively, intuitively and finely reveal the evolution of the microstructure (especially ice crystals) inside the litchi during the preservation process.

[0005] Therefore, there is an urgent need for a technical means to effectively observe the formation of ice crystals inside litchi during ice temperature preservation. Summary of the Invention

[0006] To overcome the problems existing in the related art, one of the objectives of the present invention is to provide a method for obtaining an image of the ice crystal structure of lychee flesh under a low-frequency electric field. The method for obtaining an image of the ice crystal structure of lychee flesh under a low-frequency electric field obtains k-space data inside the lychee by combining magnetic resonance imaging (MRI) technology, and uses a specifically optimized convolutional neural network (CNN) model to process and reconstruct the image space data after Fourier transformation to generate a high-definition ice crystal structure image, thereby overcoming the problem in the prior art of lacking effective, non-destructive, intuitive and sophisticated technical means to observe and analyze the formation, distribution and evolution of microscopic ice crystals inside lychee under ice temperature preservation conditions.

[0007] A method for obtaining an image of ice crystal structure of lychee flesh under a low-frequency electric field, comprising:

[0008] Magnetic resonance imaging (MRI) was performed on litchi fruits that had been treated with a low-frequency electric field and stored at ice temperature to obtain k-space data of their flesh quality.

[0009] Performing Fourier transform on the k-space data to obtain image space data containing amplitude information and phase information;

[0010] The amplitude information and phase information of the image spatial data are input as dual-channel inputs to a pre-trained convolutional neural network model to obtain an ice crystal structure image of lychee flesh with enhanced ice crystal structure clarity;

[0011] Among them, the convolutional neural network model includes a convolutional attention module, which is used to enhance the convolutional neural network model's perception ability of the characteristic channels and spatial positions of the litchi ice crystal area.

[0012] By combining k-space data acquired from magnetic resonance imaging (MRI) scans with a specific convolutional neural network model that includes a convolutional attention module, and processing dual-channel data containing both amplitude and phase information, this method can non-destructively capture microstructural images of the lychee flesh. Compared to traditional methods, this method leverages the powerful image reconstruction capabilities of deep learning models and the ability of attention mechanisms to focus on key features (ice crystal regions), significantly improving the clarity and detail of ice crystal structures. This overcomes the difficulty of traditional imaging techniques in accurately observing the tiny ice crystal structures within lychees stored at ice temperatures, providing a high-quality imaging foundation for in-depth research into the mechanisms of ice crystal formation.

[0013] Furthermore, the pre-trained convolutional neural network model is obtained through a training process comprising the following steps:

[0014] Obtaining a litchi nuclear magnetic resonance image dataset for training, and marking litchi flesh ice crystal regions in the litchi nuclear magnetic resonance image dataset;

[0015] Dividing the litchi nuclear magnetic resonance image dataset into a training set and a validation set;

[0016] The convolutional neural network model is trained using the training set, and the convolutional neural network model is verified using the verification set.

[0017] Specifically, the model was trained using a dataset of litchi MRI images, specifically acquired and manually annotated with ice crystal regions. This split into training and validation sets ensured that the convolutional neural network model was optimized for the specific task of identifying and enhancing litchi ice crystal structure. Compared to models trained using general image processing models or those not trained on specific data, the resulting model exhibited greater specificity and accuracy, more effectively identifying and reconstructing the complex features of litchi ice crystals, thereby improving the reliability and detail fidelity of the final output image.

[0018] Furthermore, in the step of training the convolutional neural network model using the training set, a composite loss function is used to optimize the model.

[0019] A composite loss function is used for optimization during model training, allowing the model to simultaneously consider multiple evaluation metrics (such as pixel-level accuracy, area overlap, and edge clarity) during learning. Compared to using only a single loss function (such as focusing only on pixel accuracy or area overlap), the composite loss function can guide the model to achieve a more balanced and comprehensive optimization goal, helping to produce images that perform better across multiple key dimensions, particularly those crucial for ice crystal morphology and boundary analysis.

[0020] Furthermore, the composite loss function is a weighted combination of a binary cross entropy loss term, a region overlap loss term, and an edge loss term.

[0021] The composite loss function is a weighted combination of binary cross-entropy loss (focusing on pixel classification accuracy), region overlap loss (focusing on the consistency between the predicted and true regions, and sensitive to small objects), and edge loss (focusing on boundary clarity). This allows for more precise optimization of the core requirements of ice crystal image reconstruction. This combination is particularly beneficial for processing the characteristics of ice crystal structure (small size and clearly defined boundaries), ensuring that the model not only roughly locates ice crystals but also accurately segments their regions and sharpens their edges, resulting in images that are more valuable for subsequent analysis.

[0022] Furthermore, the edge loss term is calculated as follows:

[0023] Applying an edge detection operator to the predicted ice crystal structure image output by the convolutional neural network model and the corresponding true annotated ice crystal structure image to obtain a predicted edge image and a true edge image respectively;

[0024] The mean square error between the predicted edge map and the true edge map is calculated to obtain the value of the edge loss term.

[0025] By applying an edge detection operator to the model's output predictions and the ground-truth labeled images and calculating the mean squared error between the edge maps to define the edge loss term, the model optimization focus is directly directed to "edge consistency." Compared to losses that rely solely on pixel values ​​or area overlap, this loss calculation method, which explicitly targets edges, more effectively forces the model to learn and reproduce the precise outlines and boundary details of ice crystal structures, significantly improving the clarity and continuity of ice crystal edges in the final image.

[0026] Furthermore, the edge detection operator is a Sobel operator.

[0027] The Sobel operator is sensitive to gradient changes in images and can effectively extract edge information of ice crystal structures. Using this specific operator not only ensures the effectiveness of edge loss calculation, but also improves the feasibility of the method and the reproducibility of the results.

[0028] Furthermore, the weight coefficient of the composite loss function is dynamically adjusted through a particle swarm optimization algorithm.

[0029] The particle swarm optimization (PSO) algorithm dynamically adjusts the weight coefficients of each loss term in the composite loss function, overcoming the difficulties of manually setting weights to achieve optimal results and the potential need for different weightings at different training stages. PSO adaptively searches for the optimal weight combination, enabling the model training process to better balance the contributions of different loss terms. This potentially achieves better overall optimization results than fixed weights, further improving model performance and final image quality.

[0030] Furthermore, the particle swarm optimization algorithm is based on a preset speed update formula and a position update formula, takes minimizing the weighted combination value of the composite loss function as the optimization goal, and iteratively searches for the optimal weight coefficient combination.

[0031] Compared with simple grid search or random search, the global optimization strategy based on swarm intelligence can more efficiently and intelligently find a combination of weight coefficients close to the global optimal state, ensuring the effectiveness and scientificity of weight optimization and helping the model to fully converge to a better state.

[0032] Furthermore, the convolutional attention module is a CBAM module, and the CBAM module sequentially includes a channel attention submodule and a spatial attention submodule;

[0033] The channel attention submodule is used to learn the importance weights of different feature channels in the litchi ice crystal image;

[0034] The spatial attention submodule is used to learn the importance weights of different spatial positions in the litchi ice crystal image.

[0035] A second object of the present invention is to provide a system for acquiring an image of the microscopic ice crystal structure of lychee flesh, which is used to perform the above-mentioned method for acquiring an image of the ice crystal structure of lychee flesh under a low-frequency electric field. The system comprises:

[0036] A magnetic resonance scanning unit is used to perform magnetic resonance scanning on lychees that have been treated with a low-frequency electric field and stored at ice temperature to obtain k-space data of the lychee flesh quality;

[0037] a data processing unit, configured to perform Fourier transform on the k-space data to obtain image space data containing amplitude information and phase information;

[0038] An image reconstruction unit includes a pre-trained convolutional neural network model, wherein the convolutional neural network model includes a convolutional attention module, and the convolutional neural network model is used to receive the amplitude information and phase information from the data processing unit as dual-channel input; the pre-trained convolutional neural network model is used to process the dual-channel input to obtain a lychee flesh ice crystal structure image with enhanced ice crystal structure clarity.

[0039] The beneficial effects of the present invention are:

[0040] The present invention provides a method for acquiring images of the ice crystal structure of lychee flesh under a low-frequency electric field. This method applies magnetic resonance imaging (MRI) technology to lychees stored at ice temperatures in a low-frequency electric field to obtain k-space data. This data is then Fourier transformed to obtain image space data containing amplitude and phase information. This dual-channel information is then input into a convolutional neural network (CNN) model containing a convolutional attention module for reconstruction. This method enables high-quality, non-destructive imaging of the microscopic ice crystal structure within the lychee flesh. Compared to the limitations of traditional observation methods, which have difficulty in deeply and precisely revealing the ice crystal morphology within tissues, this method utilizes the penetrating power of MRI (magnetic resonance imaging) to acquire k-space data of lychee flesh. By utilizing dual-channel information and an attention mechanism to enhance the perception of key features, the method significantly improves the clarity, edge sharpness, and detail of the ice crystal structure. This method provides unprecedented intuitive and reliable imaging evidence for accurately studying the microscopic mechanisms of ice crystal formation, distribution, and interaction with tissue during lychee preservation at ice temperatures. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1It is a schematic diagram of the method for obtaining the ice crystal structure image of lychee flesh under a low-frequency electric field provided in this application. DETAILED DESCRIPTION

[0042] The preferred embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although preferred embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided to make the present invention more thorough and complete and to fully convey the scope of the present invention to those skilled in the art.

[0043] Example

[0044] like Figure 1 As shown, this embodiment provides a method and system for obtaining an image of ice crystal structure of lychee flesh under a low-frequency electric field. The method for obtaining an image of ice crystal structure of lychee flesh under a low-frequency electric field comprises the following steps:

[0045] Step S1: Sample preparation and pretreatment

[0046] 200 fresh commercially available "Feizixiao" lychees were selected based on criteria: uniform fruit size (each weighing approximately 20-25g), bright red or slightly yellow-green color, and intact peel free of mechanical damage and pests. Eligible lychees were numbered and labeled. The labeled lychees were then placed in a cold storage cabinet set at -1°C, treated with a low-frequency AC electric field, and stored for a period of time.

[0047] Step S2: Magnetic resonance scanning and data acquisition

[0048] Randomly sample the litchi fruit after processing in step S1 and place it in the scanning area of ​​a magnetic resonance imaging (MRI) device. Use an appropriate pulse sequence to scan the litchi flesh area and obtain its two-dimensional k-space (frequency domain) data. The k-space data is a complex matrix, as follows:

[0049]

[0050] Where u and v represent the spatial frequency coordinates of ice crystals captured by the NMR device during icy preservation of lychees, representing the vertical and horizontal frequency component indices, respectively. The value of K(u,v) represents the signal amplitude and phase at that frequency component.

[0051] Step S3: Fourier transform and initial image generation

[0052] Apply a two-dimensional discrete Fourier transform to the k-space data matrix K(u,v) obtained in step S2 to convert it from the frequency domain to the image space domain to obtain the initial image data of the litchi. The two-dimensional discrete Fourier transform formula is as follows:

[0053]

[0054] Where M and N are the number of rows and columns in the k-space matrix (corresponding to the image resolution), and I(x,y) is the complex-valued pixel at the coordinate (x,y) in the image space domain. From the complex image data I(x,y), the amplitude information |I(x,y)| and the phase information ∠I(x,y) are separated. The resulting |I(x,y)| can be considered a preliminary grayscale image, but the ice crystal structure may not be clear enough.

[0055] Step S4: Preparation of training data (this step is completed before the model is applied)

[0056] To train the subsequent convolutional neural network model, a dataset of labeled litchi nuclear magnetic resonance images is required. A batch (e.g., 200) of icy-preserved litchi nuclear magnetic resonance images are acquired using an MRI device (these can be the amplitude images generated in step S3, or specially acquired diffusion-weighted imaging (DWI) images, the latter of which are more conducive to manual identification of ice crystals). Experienced technicians manually or semi-automatically annotate the ice crystal areas in the litchi flesh on these images and generate corresponding binary mask images (ice crystal areas are 1, background is 0). These 200 sets of images and their corresponding annotated masks are randomly divided into a training set (160 groups) and a validation set (40 groups) in a ratio of 8:2.

[0057] Step S5: Convolutional neural network model construction and training

[0058] Build a convolutional neural network (CNN) model designed for image reconstruction or segmentation tasks. The key is to replace the standard attention module (if any) in the model with a Convolutional Block Attention Module (CBAM). The CBAM module consists of a channel attention submodule and a spatial attention submodule.

[0059] The attention mechanism of traditional CNN models typically focuses only on the spatial or channel dimensions of the image, resulting in limited granularity. This can easily blur the image of small ice crystals and lacks the ability to focus on features in tiny regions. Therefore, the CBAM (Channel-Attention Module) is introduced to enhance the ability to resolve the boundaries of small ice crystals in MRI images by combining information from both channel and spatial dimensions. This allows the model to automatically focus on the tiny edges and weak texture structures of ice crystal tissue, thereby improving model performance and accuracy.

[0060] The spatial domain information of the image, obtained using the Fourier transform formula, is used to construct the CNN input feature map. Considering that the edges, texture, and structural details of litchi ice crystals rely more heavily on phase information, a dual-channel input, including amplitude and phase information, is used to more comprehensively capture image information, thereby improving the model's ability to recognize the edges, texture, and structural features of ice crystal regions. The formula is as follows:

[0061] F=Input(x,y)=[|I(x,y)|,∠I(x,y)]

[0062] In the CBAM module, the input feature map first passes through the channel attention mechanism to filter out the important feature channels of the ice crystal structure area that are critical for the preservation of lychees at ice temperature. The channel attention mechanism uses global pooling to compress spatial information to obtain the statistical characteristics of each channel. The attention weights of these channels are learned through the MLP (Multi-layer Perceptron), allowing the network to focus on the feature channels that are important for the preservation of lychees at ice temperature. The formula is as follows:

[0063] M c (F)=σ(MLP(AvgPool(F))+MLP(MaxPool(F)))

[0064] Among them, σ represents a sigmoid activation function, which normalizes the value; AvgPool(F) represents the average pooling value of the input feature map, and the formula is as follows:

[0065]

[0066] Where H represents the height of the feature map, that is, the number of pixels in the vertical direction of the image; W represents the width of the feature map, that is, the number of pixels in the horizontal direction of the image; c represents the number of channels in the feature map;

[0067] MaxPool(F) represents the maximum value of all pixels in the input feature map channel. The formula is as follows:

[0068] MaxPool(F)=max x,y F(c,x,y)

[0069] The input feature map is processed by the CBAM channel attention module to obtain F', the formula is as follows:

[0070] F'=M c (F)⊙F

[0071] Among them, ⊙ represents element-by-element multiplication; semantic features are extracted through the CNN model and the prediction map is output;

[0072] Fdeep =CNN(F')

[0073] Among them, CNN represents a complete convolutional neural network structure. deep Represents the extracted deep semantic feature map;

[0074] The prediction graph is then output through convolution transformation and activation function as follows:

[0075]

[0076] Among them, W 1×1 It is a 1×1 convolution kernel, which is used to transform the dual-channel F deep The mapping is single-channel; * represents a convolution operation; b represents a scalar or a bias map with the same size as the output map; Represents the intermediate prediction value after convolution; σ represents the Sigmoid function; P(x,y) represents the final prediction map.

[0077] The amplitude information and phase information obtained in step S3 are provided to the CNN model as a dual-channel input feature map.

[0078] The CBAM-CNN model is trained using the training set (160 images and their annotations) prepared in step S4. During training, a composite loss function is used for optimization. This composite loss function is a weighted combination of binary cross entropy loss (BCE), Dice loss (also known as region overlap loss), and edge loss (EdgeLoss):

[0079]

[0080] Among them, Loss total is the composite loss function value.

[0081] Binary cross entropy (BCE) calculates the cross entropy between the pixel probability predicted by the model and the true label (0 or 1). The formula is as follows:

[0082]

[0083] Where N represents the total number of pixels in the image N = H × W, Y(x, y) represents the pixel value in the real NMR image, and P(x, y) represents the pixel point in the predicted image.

[0084] The DiceLoss region overlap loss is calculated by calculating the degree of overlap between the predicted ice crystal region and the true labeled region (Dice coefficient), which helps improve the model's sensitivity to the formation of small ice crystals in litchi flesh during ice temperature preservation. The formula is as follows:

[0085]

[0086] DiceLoss = 1-Dice

[0087] Edge loss EdgeLoss facilitates the model to measure the edge relationship between the predicted image and the real NMR image of the ice crystal structure, and is calculated as follows:

[0088] Use Sobel convolution to extract the edge of ice crystal tissue and define the convolution kernel G in the horizontal and vertical directions x and G y ;

[0089]

[0090] Convolution operations are performed on the predicted image and the real NMR image to obtain E x and E y ;

[0091]

[0092] Calculate the mean squared error (MSE) between edge maps

[0093]

[0094] The model is trained using the training set. During the training process, the particle swarm optimization (PSO) algorithm is used to dynamically adjust the weight coefficient λ of the loss function. 1 ,λ 2 and λ 3 .make The arithmetic function λ representing the k-th generation example 1 ,λ 2 and λ 3 Weight combination, the particle swarm optimization algorithm includes speed update and position update formulas. The speed update formula is as follows:

[0095]

[0096] Where c1 and c2 represent acceleration constants; r1 and r2 represent random numbers between 0 and 1; w is the inertia weight; Represents the comprehensive loss function value corresponding to the k-th generation particle i; Represents the minimum comprehensive loss function value experienced by the k-th generation particle; Represents the minimum comprehensive loss function value of the entire population in the kth generation;

[0097] The position update formula is as follows:

[0098]

[0099] represents the position of the i-th particle in the (k+1)th generation; represents the position of the i-th particle in the k-th generation; represents the velocity of the i-th particle in the (k+1)th generation.

[0100] The three loss functions and corresponding weights are combined into a comprehensive loss function as the objective function R of the particle swarm optimization algorithm. The calculation formula is as follows:

[0101]

[0102] During the training process, the validation set (40 groups) prepared in step S4 is regularly used to evaluate model performance to prevent overfitting and to adjust hyperparameters or determine the optimal number of training rounds. After training is completed, the optimized model weights are saved.

[0103] Step S6: Apply the trained model to reconstruct the image

[0104] The new dual-channel (amplitude and phase) initial image data obtained through steps S1-S3 is fed into the CBAM-CNN model trained in step S5. Through forward propagation, the model leverages its learned feature extraction, attention focus, and reconstruction capabilities to output a final image of the ice crystal structure of the lychee flesh. Compared to the initial amplitude image generated in step S3, this image features clearer ice crystal structure outlines, richer details, and less background noise, resulting in significantly enhanced clarity.

[0105] This embodiment also provides a system for obtaining an image of the microscopic ice crystal structure of lychee flesh, the system comprising:

[0106] Magnetic resonance scanning unit: for example, a low-field nuclear magnetic resonance apparatus, used to perform the scanning operation in step S2 and acquire k-space data.

[0107] Data processing unit: For example, a computer equipped with appropriate computing software (such as MATLAB or Python scientific computing libraries). This unit is used to perform the two-dimensional inverse discrete Fourier transform in step S3 to calculate the initial image space data containing amplitude and phase information from the k-space data.

[0108] The image reconstruction unit can be the same computer as the data processing unit or a dedicated image processing workstation. This unit is loaded with the convolutional neural network model software, including the CBAM module, trained in step S5. It receives the dual-channel (amplitude and phase) data from the data processing unit as input, performs the model forward propagation calculations in step S6, and ultimately outputs an image of the lychee flesh ice crystal structure with enhanced ice crystal clarity.

[0109] This implementation provides a method and system for acquiring images of ice crystal structure in lychee flesh under a low-frequency electric field. MRI technology is used to achieve non-destructive scanning of the internal structure of lychee, and to obtain the original k-space data required for subsequent analysis. The amplitude and phase information after Fourier transformation are used to form a dual-channel input, fully utilizing the advantages of phase information in reflecting edge and texture details, providing the subsequent CNN model with more comprehensive image information than a single amplitude channel, which helps to improve the model's ability to recognize ice crystal structural features. A CNN model containing the CBAM attention mechanism is used. Through its ability to combine channel and spatial dimensions of attention, the model's ability to focus on and resolve small, weakly textured ice crystal areas and their boundaries in lychee MRI images is effectively enhanced, overcoming the limitations of traditional attention mechanisms. The model is trained using a litchi nuclear magnetic resonance image dataset specifically labeled for litchi ice crystals, ensuring its high adaptability and accuracy for specific tasks. A composite loss function combining binary cross entropy (pixel-level accuracy), Dice loss (regional overlap, sensitive to small targets), and edge loss (boundary clarity) based on the Sobel operator was used to comprehensively guide model optimization from multiple dimensions, which is particularly in line with the needs of small particle recognition and edge sharpening in ice crystal image reconstruction. The particle swarm optimization (PSO) algorithm was introduced to dynamically and adaptively adjust the weight coefficients of the composite loss function, overcoming the difficulties and suboptimality of manually setting or fixing weights. A better weight balance point was found through an automated optimization mechanism, thereby further improving the overall performance of the model and the final image quality. Compared with the original MRI image, the final output of the lychee flesh ice crystal structure image has been significantly enhanced in terms of the clarity of the ice crystal outline, the richness of the internal details, and the suppression of background noise. It provides researchers with high-quality, intuitive visual evidence and greatly facilitates in-depth research, quantitative analysis, and optimization of the ice crystal formation mechanism of lychee under specific ice temperature preservation conditions. The provided system integrates scanning, data processing and image reconstruction based on advanced models, providing a complete solution for the standardized and efficient application of this method.

[0110] Unless otherwise specifically stated, the relative arrangement, numerical expression and numerical value of the parts and steps set forth in these embodiments do not limit the scope of the application. In all examples shown and discussed here, any specific value should be interpreted as merely exemplary, rather than as a restriction. Therefore, other examples of exemplary embodiments can have different values. It should be noted that: similar reference numerals and letters represent similar items in the accompanying drawings below, and therefore, once a certain item is defined in an accompanying drawing, it does not need to be further discussed in the accompanying drawings subsequently.

[0111] In addition, it should be noted that the use of terms such as "first" and "second" for limitation is only for the convenience of distinction. Unless otherwise stated, the above terms have no special meaning and therefore cannot be understood as limiting the scope of protection of this application.

[0112] The foregoing description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be within the scope of protection of the present invention.

Claims

1. A method for obtaining an image of ice crystal structure of lychee flesh under a low-frequency electric field, characterized in that: include: Magnetic resonance imaging (MRI) was performed on litchi fruits that had been treated with a low-frequency electric field and stored at ice temperature to obtain k-space data of their flesh quality. Performing Fourier transform on the k-space data to obtain image space data containing amplitude information and phase information; The amplitude information and phase information of the image spatial data are input as dual-channel inputs to a pre-trained convolutional neural network model to obtain an ice crystal structure image of lychee flesh with enhanced ice crystal structure clarity; Among them, the convolutional neural network model includes a convolutional attention module, which is used to enhance the convolutional neural network model's perception ability of the characteristic channels and spatial positions of the litchi ice crystal area.

2. The method for acquiring an image of ice crystal structure of lychee flesh under a low-frequency electric field according to claim 1, characterized in that: The pre-trained convolutional neural network model is trained by the following steps: Obtaining a litchi nuclear magnetic resonance image dataset for training, and marking litchi flesh ice crystal regions in the litchi nuclear magnetic resonance image dataset; Dividing the litchi nuclear magnetic resonance image dataset into a training set and a validation set; The convolutional neural network model is trained using the training set, and the convolutional neural network model is verified using the verification set.

3. The method for acquiring an image of ice crystal structure of lychee flesh under a low-frequency electric field according to claim 2, characterized in that: In the step of using the training set to train the convolutional neural network model, a composite loss function is used to optimize the model.

4. The method for acquiring an image of ice crystal structure of lychee flesh under a low-frequency electric field according to claim 3, characterized in that: The composite loss function is a weighted combination of a binary cross entropy loss term, a region overlap loss term, and an edge loss term.

5. The method for acquiring an image of ice crystal structure of lychee flesh under a low-frequency electric field according to claim 4, characterized in that: The edge loss term is calculated as follows: Applying an edge detection operator to the predicted ice crystal structure image output by the convolutional neural network model and the corresponding true annotated ice crystal structure image to obtain a predicted edge image and a true edge image respectively; The mean square error between the predicted edge map and the true edge map is calculated to obtain the value of the edge loss term.

6. The method for acquiring an image of ice crystal structure of lychee flesh under a low-frequency electric field according to claim 5, characterized in that: The edge detection operator is a Sobel operator.

7. The method for acquiring an image of ice crystal structure of lychee flesh under a low-frequency electric field according to claim 4, characterized in that: The weight coefficient of the composite loss function is dynamically adjusted through a particle swarm optimization algorithm.

8. The method for acquiring an image of ice crystal structure of lychee flesh under a low-frequency electric field according to claim 7, characterized in that: The particle swarm optimization algorithm is based on a preset speed update formula and a position update formula, takes minimizing the weighted combination value of the composite loss function as the optimization goal, and iteratively searches for the optimal weight coefficient combination.

9. The method for acquiring an image of ice crystal structure of lychee flesh under a low-frequency electric field according to claim 1, characterized in that: The convolutional attention module is a CBAM module, which includes a channel attention submodule and a spatial attention submodule in sequence; The channel attention submodule is used to learn the importance weights of different feature channels in the litchi ice crystal image; The spatial attention submodule is used to learn the importance weights of different spatial positions in the litchi ice crystal image.

10. A system for obtaining images of the microscopic ice crystal structure of lychee flesh, characterized in that: The system for executing the method for acquiring an image of ice crystal structure of lychee flesh under a low-frequency electric field according to any one of claims 1 to 9 comprises: A magnetic resonance scanning unit is used to perform magnetic resonance scanning on lychees that have been treated with a low-frequency electric field and stored at ice temperature to obtain k-space data of the lychee flesh quality; a data processing unit, configured to perform Fourier transform on the k-space data to obtain image space data containing amplitude information and phase information; An image reconstruction unit includes a pre-trained convolutional neural network model, wherein the convolutional neural network model includes a convolutional attention module, and the convolutional neural network model is used to receive the amplitude information and phase information from the data processing unit as dual-channel input; the pre-trained convolutional neural network model is used to process the dual-channel input to obtain a lychee flesh ice crystal structure image with enhanced ice crystal structure clarity.

Images