An image processing system for identifying the anthocyanin content of colored potatoes

Abstract

The application discloses a kind of identification color potato anthocyanin content image processing system, by setting up image acquisition unit obtains three kinds of images of visible light, epidermis texture and spectrum, and carries out weighted analysis, enhances the representativeness and distinguish degree of image data;Illumination correction unit adjusts image compensation strategy dynamically according to ambient light intensity, significantly reduces the feature extraction deviation caused by uneven illumination;Feature selection unit fuses multi-source image data and extracts target comprehensive feature vector, fully excavates the complementary information between multi-modal data;Prediction unit generates anthocyanin content function and distribution heat map based on the target comprehensive feature vector, improves the recognition accuracy and visualization ability, overcomes the defects of traditional method sensitive to light, weak non-linear modeling ability, realizes efficient, accurate and stable identification of potato anthocyanin content.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to an image processing system for identifying the anthocyanin content of colored potatoes. Background Technology

[0002] Identifying the anthocyanin content of colored potatoes is an indispensable part of modern industry for assessing their nutritional value, variety selection, and processing applications. Anthocyanins are natural antioxidants with health benefits such as enhancing immunity, delaying aging, and protecting the cardiovascular system and vision. The types and contents of anthocyanins vary significantly among different varieties and parts of the plant, directly affecting their commercial value and functional properties. Accurate detection can guide the breeding of high-anthocyanin varieties, optimize processing techniques, and provide a scientific basis for the development of functional foods, while simultaneously meeting consumers' demands for healthy eating.

[0003] In existing technologies, due to the high diversity of potato skin texture and uneven color distribution, traditional models have difficulty accurately capturing the nonlinear relationship between color parameters and anthocyanin content, and are easily affected by light interference, leading to feature extraction deviations and a decrease in recognition accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide an image processing system for identifying the anthocyanin content of colored potatoes, in order to solve the technical problems that traditional models are difficult to accurately capture the nonlinear relationship between color parameters and anthocyanin content, and are easily affected by light interference, which leads to feature extraction deviation and a decrease in recognition accuracy.

[0005] The technical solution of this invention is implemented as follows:

[0006] An image processing system for identifying the anthocyanin content of colored potatoes includes: an image acquisition unit, an illumination correction unit, a feature selection unit, and a prediction unit;

[0007] The image acquisition unit is used to acquire visible light images, skin texture images, and spectral images of potatoes and perform weighted analysis on the images respectively. The weighted analysis includes an image weighting module, which is used to stitch the visible light images, skin texture images, and spectral images of potatoes into a feature matrix, and assign weights according to the information contribution rate of each mode of the feature matrix to generate target visible light images, target skin texture images, and target spectral images.

[0008] The illumination correction unit is used to adaptively correct the compensation strategies corresponding to the visible light image, skin texture image and spectral image of the potato according to the light intensity of the surrounding environment of the potato.

[0009] The feature selection unit is used to fuse the target visible light image, the target epidermal texture image, and the target spectral image to generate target image data, and select the target comprehensive feature vector of the target image data based on the target neural network;

[0010] The prediction unit performs mathematical calculations based on the target comprehensive feature vector to output the target content function of potato anthocyanins and distribution heatmap data.

[0011] A further technical solution is that the image acquisition unit includes:

[0012] The image acquisition module uses a visible light camera, a 3D scanner, and a near-infrared spectral imager to acquire visible light images, skin texture images, and spectral images of the potato.

[0013] The image weighting module is used to stitch together the potato visible light image, skin texture image and spectral image into a feature matrix, and to assign weights according to the information contribution rate of each mode of the feature matrix;

[0014] The data processing module is used to remove low-weight and high-weight values ​​of the information contribution rate of each mode of the feature matrix, re-integrate the remaining weight values, and establish a new stitching relationship with the potato visible light image, skin texture image and spectral image to construct a weighted covariance matrix.

[0015] The target output module is used to generate the corresponding target visible light image, target skin texture image, and target spectral image from the potato visible light image, skin texture image, and spectral image based on the weighted covariance matrix.

[0016] A further technical solution is that the image weighting module performs the following steps:

[0017] Step S11: Extract the color histogram, gray-level co-occurrence matrix features, and band reflectance curves corresponding to the visible light image, skin texture image, and spectral image of the potato, respectively.

[0018] Step S12: Input the color histogram, gray-level co-occurrence matrix features and band reflectance curve into a multi-scale feature pyramid network for modal feature interaction to generate a hybrid feature vector containing local details and global semantics.

[0019] Step S13: Calculate the feature importance score of each modality for the target detection task using a decision tree model, and generate weight coefficients by combining the sensitivity of spectral data to anthocyanin detection.

[0020] Step S14: Input the weight coefficients into the gating attention mechanism, and perform nonlinear weighting on the feature channels of different modalities through learnable attention weights.

[0021] A further technical solution is that the data processing module performs the following steps:

[0022] Step S21: Standardize all feature data and filter out data that has little impact on the results or fluctuates abnormally.

[0023] Step S22: Extract color, texture and spectral features respectively, and unify the color, texture and spectral features to the same frequency dimension through multi-scale decomposition;

[0024] Step S23: Combine the multi-source data of the same frequency dimension, calculate the correlation between the multi-source data, and assign different weights according to importance to establish data relationships.

[0025] A further technical solution is that the illumination correction unit performs the following steps:

[0026] Step S31: Simultaneously use a regular camera and a spectrometer to measure the light intensity of the environment around the potato in real time;

[0027] Step S32: Dynamically adjust the brightness of the visible light image, and extract the fine texture features of the surface and the reflection information of different color bands to initially lock in the anthocyanin features.

[0028] Step S33: Using intelligent algorithms, the anthocyanin features are amplified and noise in the image is reduced in low light or high contrast conditions.

[0029] Step S34: Based on the sensitivity of anthocyanins to specific color bands, automatically adjust the weight ratio of each band and jointly optimize the weight ratio.

[0030] A further technical solution is that step S33 specifically includes:

[0031] Step 331: Separate the illumination component and anthocyanin reflectance component of the low-light image;

[0032] Step 332: Use a loss function to constrain the structural similarity between the illumination component and the anthocyanin reflection component, generate a pre-processed image, and simultaneously remove random noise from the image;

[0033] Step 333: Based on the LLCNN architecture, perform multi-channel analysis on the processed image to generate an enhanced image that highlights anthocyanin features;

[0034] Step 334: Enhance the anthocyanin-related features in the enhanced image and filter out the residual noise again.

[0035] A further technical solution is that the feature selection unit performs the following steps:

[0036] Step S41: Adjust the target visible light image, target skin texture image and target spectral image to a uniform clarity and color range, eliminate the influence of light, and extract the common features of the target visible light image, target skin texture image and target spectral image;

[0037] Step S42: Analyze the color changes in visible light, extract the surface texture features, and calculate the band data corresponding to the color depth in the spectral image.

[0038] Step S43: Assign importance based on the correlation degree of target features in each of the target visible light image, target skin texture image and target spectral image, and remove duplicate or irrelevant information;

[0039] Step S44: The key information extracted from the target visible light image, target skin texture image and target spectral image is spliced ​​into a complete data package, retaining the core data that best represents the characteristics of the target;

[0040] Step S45: Input the core data into the target neural network, adjust the parameters of the target neural network to find the best feature combination, and verify whether the feature combination accurately judges different target features;

[0041] Step S46: Output the target features and generate a comprehensive feature vector.

[0042] A further technical solution is that step S45 specifically includes:

[0043] Step S451: Input the core data into the preset model;

[0044] Step S452: By automatically adjusting the internal weights and biases of the preset model, try different combinations of the features;

[0045] Step S453: Use the validation set to monitor the discriminative power of the feature combination in real time and record the optimal parameter configuration;

[0046] Step S454: Use a test set to check whether the preset model can accurately identify different target features;

[0047] Step S455: If the recognition accuracy does not meet the standard, return to step S452 to readjust the parameters until the feature combination achieves the expected effect.

[0048] A further technical solution is that the prediction unit performs the following steps:

[0049] Step S51: Input the target comprehensive feature vector into the preset analysis model;

[0050] Step S52: Automatically calculate and generate the objective function for anthocyanin content using the preset analysis model;

[0051] Step S53: Verify the actual detection results of the sample and adjust the parameters of the objective function;

[0052] Step S54: Calculate the anthocyanin concentration value of each pixel according to the objective function, and generate corresponding distribution heatmap data based on color intensity;

[0053] Step S55: Integrate the objective function and heatmap data into an output result file.

[0054] A further technical solution is that step S52 specifically includes:

[0055] Step S521: Use the target integrated feature vector as input data;

[0056] Step S522: Automatically identify the mathematical relationship between the input data and anthocyanin content using a clustering algorithm;

[0057] Step S523: Iteratively optimize the model parameters using known sample data to minimize the error between the predicted result and the actual value;

[0058] Step S524: Output the final formula for calculating anthocyanin content;

[0059] Step S525: Encapsulate the calculation formula into a directly callable function module for subsequent batch calculations or real-time predictions.

[0060] The beneficial effects of this invention are as follows:

[0061] By setting up an image acquisition unit to acquire three types of images—visible light, epidermal texture, and spectrum—and performing weighted analysis, the representativeness and discriminative power of the image data are enhanced. The illumination correction unit dynamically adjusts the image compensation strategy according to the ambient light intensity, significantly reducing the feature extraction bias caused by uneven illumination. The feature selection unit fuses multi-source image data and extracts the target comprehensive feature vector, fully exploring the complementary information between multimodal data. The prediction unit generates anthocyanin content function and distribution heatmap based on the target comprehensive feature vector, improving recognition accuracy and visualization capabilities. This overcomes the shortcomings of traditional methods, such as sensitivity to light and weak nonlinear modeling ability, and achieves efficient, accurate, and stable identification of potato anthocyanin content. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a schematic diagram of the structure of an image processing system for identifying the anthocyanin content of colored potatoes according to the present invention;

[0064] Figure 2 This is a schematic diagram of the image acquisition unit structure in an image processing system for identifying anthocyanin content in colored potatoes according to the present invention.

[0065] Figure 3 A flowchart illustrating the execution steps of an image weighting module in an image processing system for identifying anthocyanin content in colored potatoes according to the present invention;

[0066] Figure 4 A flowchart illustrating the execution steps of a data processing module in an image processing system for identifying anthocyanin content in colored potatoes according to the present invention.

[0067] Figure 5 A flowchart of the execution steps of the illumination correction unit in an image processing system for identifying anthocyanin content in colored potatoes according to the present invention;

[0068] Figure 6 A flowchart of the execution steps of the feature selection unit in an image processing system for identifying anthocyanin content in colored potatoes according to the present invention;

[0069] Figure 7 A flowchart of the execution steps of the prediction unit in an image processing system for identifying anthocyanin content in colored potatoes according to the present invention. Detailed Implementation

[0070] To better understand the technical content of this invention, specific embodiments are provided below, and the invention will be further described in conjunction with the accompanying drawings.

[0071] See Figures 1 to 7This invention provides an image processing system for identifying the anthocyanin content of colored potatoes, comprising: an image acquisition unit, an illumination correction unit, a feature selection unit, and a prediction unit; the image acquisition unit is used to acquire visible light images, skin texture images, and spectral images of potatoes and perform weighted analysis on each image, wherein the weighted analysis includes an image weighting module, used to concatenate the visible light images, skin texture images, and spectral images of potatoes into a feature matrix, and assign weights according to the information contribution rate of each mode of the feature matrix to generate a target visible light image, a target skin texture image, and a target spectral image; the illumination correction unit is used to adaptively correct the compensation strategies corresponding to the visible light images, skin texture images, and spectral images of potatoes according to the light intensity of the surrounding environment of the potato; the feature selection unit is used to fuse the target visible light images, target skin texture images, and target spectral images to generate target image data, and select a target comprehensive feature vector of the target image data based on a target neural network; the prediction unit performs mathematical calculations based on the target comprehensive feature vector to output the target anthocyanin content function and distribution heatmap data of potatoes.

[0072] This invention utilizes a sequentially connected image acquisition unit, illumination correction unit, feature selection unit, and prediction unit to work collaboratively and achieve accurate identification of anthocyanin content in colored potatoes. The image acquisition unit simultaneously acquires visible light, skin texture, and spectral images, and enhances anthocyanin-related bands and surface texture through weighted analysis. The illumination correction unit uses a dynamic adjustment compensation strategy based on ambient light intensity to eliminate image distortion caused by uneven illumination. The feature selection unit fuses multi-source image data and uses a neural network to extract a comprehensive feature vector containing color, texture, and spectrum, uncovering nonlinear correlations. The prediction unit outputs anthocyanin content function and distribution heatmap through mathematical modeling, enabling quantitative analysis and visualization.

[0073] It is worth noting that multimodal data fusion compensates for the limitations of a single image and enhances feature representation capabilities; adaptive illumination correction technology effectively reduces environmental interference and ensures data stability; neural network-driven feature extraction and modeling are significantly superior to traditional linear methods, accurately capturing the complex relationship between anthocyanin content and color parameters, ultimately achieving high-precision, low-error anthocyanin detection.

[0074] In this embodiment of the invention, an image acquisition unit acquires three types of images—visible light, epidermal texture, and spectrum—and performs weighted analysis, enhancing the representativeness and discriminative power of the image data. An illumination correction unit dynamically adjusts the image compensation strategy based on ambient light intensity, significantly reducing feature extraction bias caused by uneven illumination. A feature selection unit fuses multi-source image data and extracts a comprehensive feature vector of the target, fully leveraging the complementary information between multimodal data. A prediction unit generates an anthocyanin content function and distribution heatmap based on this comprehensive feature vector, improving recognition accuracy and visualization capabilities. This overcomes the shortcomings of traditional methods, such as sensitivity to light and weak nonlinear modeling ability, achieving efficient, accurate, and stable identification of potato anthocyanin content.

[0075] Preferably, the image acquisition unit includes: an image acquisition module, which uses a visible light camera, a 3D scanner, and a near-infrared spectral imager to acquire visible light images, skin texture images, and spectral images of potatoes; an image weighting module, which stitches the visible light images, skin texture images, and spectral images of potatoes into a feature matrix and assigns weights according to the information contribution rate of each mode in the feature matrix; a data processing module, which removes low-weight and high-weight values ​​of the information contribution rate of each mode in the feature matrix, re-integrates the remaining weight values, and establishes a new stitching relationship with the visible light images, skin texture images, and spectral images of potatoes to construct a weighted covariance matrix; and a target output module, which generates corresponding target visible light images, target skin texture images, and target spectral images from the visible light images and skin texture images of potatoes according to the weighted covariance matrix.

[0076] In this embodiment, the information contribution rate can be that the spectral data is more sensitive to anthocyanins than visible light; the key feature can be the spectral response of a specific band; low weight can be noise or irrelevant features; and high weight can be the risk of overfitting.

[0077] The image acquisition module uses a visible light camera, a 3D scanner, and a near-infrared spectral imager to simultaneously acquire visible light images of potato characteristics (color), microstructure (skin texture), and wavelength characteristic spectral images, achieving complementary coverage of multi-dimensional data. The image weighting module dynamically allocates weights by analyzing the information contribution rate of each modality in the feature matrix, highlighting key features. The data processing module removes low-weight and high-weight values, retains robust features with medium weights, and re-integrates multi-modal data based on the remaining weights to construct a weighted covariance matrix, effectively eliminating redundant information interference. The target output module uses this covariance matrix to reconstruct the visible light and texture images, generating a target visible light image (enhancing color contrast), a target skin texture image (refining surface structure), and a target spectral image (extracting key bands), ultimately providing high-quality input for subsequent modeling.

[0078] By dynamically weighting and fusing multimodal data, the problem of insufficient modeling ability of the coupling relationship between color and texture in traditional single image processing methods is solved. At the same time, by constructing a weighted covariance matrix, the robustness of the model to illumination changes and noise is significantly improved, achieving high accuracy and anti-interference ability for anthocyanin content identification.

[0079] Preferably, the image weighting module performs the following steps:

[0080] Step S11: Extract the color histogram, gray-level co-occurrence matrix features, and band reflectance curves corresponding to the visible light image, skin texture image, and spectral image of the potato, respectively.

[0081] Step S12: Input the color histogram, gray-level co-occurrence matrix features and band reflectance curve into the multi-scale feature pyramid network for modal feature interaction, and generate a hybrid feature vector containing local details and global semantics.

[0082] Step S13: Calculate the feature importance score of each modality for the target detection task using the decision tree model, and generate weight coefficients by combining the sensitivity of spectral data to anthocyanin detection.

[0083] Step S14: Input the weight coefficients into the gating attention mechanism, and perform nonlinear weighting on the feature channels of different modalities through learnable attention weights.

[0084] In this embodiment, the color histogram can represent the macroscopic color distribution; the gray-level co-occurrence matrix feature can capture the microscopic texture pattern; the band reflectance curve can quantify the spectral response; the target detection task can be the sensitivity of spectral data to anthocyanins; and the nonlinear weighting of feature channels can be achieved by adjusting the weights of different colors, textures, and spectra in the image, so that the model can focus more on key information and weaken interference, thereby improving the accuracy and efficiency of identifying anthocyanin content.

[0085] By extracting color histograms, gray-level co-occurrence matrix features, and band reflectance curves, a multi-dimensional feature foundation is provided for subsequent fusion. A multi-scale feature pyramid network is used to break down information barriers between modalities. Through top-down semantic guidance and lateral connection detail enhancement, the color features of visible light, the spatial correlation of texture, and the wavelength characteristics of the spectrum achieve deep interaction at the local detail and global semantic levels, generating a hybrid feature vector. A decision tree model quantifies the contribution of each modality to the target detection task; that is, spectral data scores higher due to its high sensitivity to anthocyanins. Weighting coefficients are set based on domain knowledge to avoid the subjectivity of traditional manual weighting. A gated attention mechanism is introduced, using learnable attention weights to non-linearly weight the feature channels of different modalities, dynamically suppressing illumination noise and enhancing the anthocyanin band.

[0086] This module overcomes the limitations of traditional methods that rely on fixed weights (such as linear weighting or simple concatenation) for multimodal feature fusion. It combines multi-scale feature interaction, dynamic weight generation, and adaptive attention mechanism to achieve intelligent and adaptive feature fusion.

[0087] Preferably, the data processing module performs the following steps:

[0088] Step S21: Standardize all feature data and filter out data that has little impact on the results or fluctuates abnormally.

[0089] Step S22: Extract color, texture and spectral features respectively, and unify color, texture and spectral features to the same frequency dimension through multi-scale decomposition;

[0090] Step S23: Combine multi-source data of the same frequency dimension, calculate the correlation between multi-source data, and assign different weights according to importance to establish data relationships.

[0091] In this embodiment, Z-score normalization is performed on color, texture, and spectral features to eliminate dimensional differences between different modalities. Simultaneously, statistical thresholding is used to filter out pixel noise data with low contribution rates or caused by abrupt changes in illumination, ensuring the reliability of subsequent analysis. NSCT decomposition technology is used to unify macroscopic color distribution, texture microstructure, and spectral wavelength response features into the low-frequency band to preserve the overall trend of color and spectrum, while the high-frequency band preserves texture details, preventing fusion bias caused by inconsistencies in the scale of multimodal features. By calculating the correlation between multi-source data and dynamically assigning weights—that is, spectral features with higher sensitivity to anthocyanins are given higher weights—data relationships are constructed to achieve collaborative modeling between features.

[0092] Preferably, the illumination correction unit performs the following steps:

[0093] Step S31: Simultaneously use a regular camera and a spectrometer to measure the light intensity of the environment around the potato in real time;

[0094] Step S32: Dynamically adjust the brightness of the visible light image, and extract the fine texture features of the surface and the reflection information of different color bands to initially lock in the anthocyanin features.

[0095] Step S33: Using intelligent algorithms, the anthocyanin features are amplified and noise in the image is reduced in low light or high contrast conditions.

[0096] Step S34: Based on the sensitivity of anthocyanins to specific color bands, automatically adjust the weight ratio of each band and jointly optimize the weight ratio.

[0097] In this embodiment, the intensity of visible light is captured in real time by a regular camera, and the wavelength distribution of ambient light is quantified by a spectrometer for joint measurement, enabling the differentiation of color temperature differences between natural light and artificial light sources. Adaptive histogram equalization adjustment, combined with the extraction of fine textures and color band reflection information, allows for the initial localization of anthocyanin features in the visible light image. A U-Net-based attention mechanism algorithm enhances anthocyanin features and suppresses noise in scenes with alternating light and shadow contrasts and strong light, avoiding feature loss due to sudden changes in illumination. Based on the high sensitivity of anthocyanins to the 520-560nm wavelength band, an adaptive weight allocation algorithm dynamically adjusts the contribution of each color channel (e.g., enhancing the blue-green band and suppressing the red band), achieving precise weighted optimization of spectral features.

[0098] By combining multi-sensor environmental perception, dynamic feature enhancement, intelligent noise reduction and band adaptive weighting, the problem of anthocyanin features being easily distorted and strong noise interference caused by light color temperature fluctuations of ±2000K due to the temperature difference between day and night is avoided.

[0099] Preferably, step S33 specifically includes:

[0100] Step 331: Separate the illumination component and anthocyanin reflectance component of the low-light image;

[0101] Step 332: Use the loss function to constrain the structural similarity between the illumination component and the anthocyanin reflection component to generate a pre-processed image, and simultaneously remove random noise from the image;

[0102] Step 333: Based on the LLCNN architecture, perform multi-channel analysis on the processed image to generate an enhanced image that highlights anthocyanin features;

[0103] Step 334: Enhance the anthocyanin-related features in the image and filter out any remaining noise again.

[0104] In this embodiment, the illumination component and anthocyanin reflectance component of a low-light image are separated based on Retinex theory. A structural similarity loss function is introduced to suppress artifacts caused by abrupt changes in illumination. Simultaneously, nonlocal mean filtering is used to remove random noise, overcoming the noise amplification defect caused by division operations. Parallel processing of RGB and spectral channel analysis is performed based on a lightweight convolutional neural network. A learnable channel attention mechanism dynamically enhances the reflectance of the anthocyanin purple region, reducing computational load. Combining frequency domain filtering and spatial domain adaptive denoising, secondary denoising is applied to the enhanced image, avoiding the coupling interference between residual noise and anthocyanin features. This achieves accurate extraction of anthocyanin features under low-light conditions.

[0105] Preferably, the feature selection unit performs the following steps:

[0106] Step S41: Adjust the target visible light image, target skin texture image and target spectral image to a uniform clarity and color range, eliminate the influence of light, and extract the common features of the target visible light image, target skin texture image and target spectral image;

[0107] Step S42: Analyze the color changes in visible light, extract the surface texture features, and calculate the band data corresponding to the color depth in the spectral image.

[0108] Step S43: Assign importance based on the correlation between the target features in each of the target visible light image, target skin texture image, and target spectral image, and remove duplicate or irrelevant information;

[0109] Step S44: The key information extracted from the target visible light image, target skin texture image and target spectral image is spliced ​​into a complete data package, retaining the core data that best represents the target features;

[0110] Step S45: Input the core data into the target neural network, adjust the parameters of the target neural network to find the best feature combination, and verify whether the feature combination can accurately judge different target features;

[0111] Step S46: Output the target features and generate a comprehensive feature vector.

[0112] In this embodiment, illumination interference is eliminated through standardization, achieving alignment of visible light, skin texture, and spectral images in terms of sharpness and color range. Color gradient, surface unevenness, and spectral reflectance are extracted to construct a multi-dimensional feature description system. Based on the mutual information matrix, the strong correlation between color and spectrum in the purple region is dynamically calculated. By retaining features with information entropy > 0.8 and removing duplicate information, an adaptive stitching strategy is adopted to preserve the coupling relationship between anthocyanin purple band reflectance and surface roughness, avoiding the dimensionality explosion caused by traditional simple stitching. Through the convolutional kernel weights of ResNet-18, the synergistic enhancement mode of high reflectance bands of anthocyanin features and specific texture features is automatically searched. The generated comprehensive feature vector can be directly input into the classification model to achieve end-to-end feature-prediction mapping, thereby avoiding the distortion of anthocyanin features and poor model generalization ability caused by illumination fluctuations of ±30%.

[0113] Preferably, step S45 specifically includes:

[0114] Step S451: Input the core data into the preset model;

[0115] Step S452: By automatically adjusting the internal weights and biases of the preset model, try different feature combinations;

[0116] Step S453: Use the validation set to monitor the discriminative power of feature combinations in real time and record the optimal parameter configuration;

[0117] Step S454: Use the test set to check whether the preset model can accurately identify different target features;

[0118] Step S455: If the recognition accuracy does not meet the standard, return to step S452 to readjust the parameters until the feature combination achieves the expected effect.

[0119] In this embodiment, the weights and biases are automatically adjusted through a preset model to try feature combinations of different color bands and texture features in synergistic patterns. The model's discriminative ability is monitored in real time using a validation set, and the optimal parameter configuration is recorded. If the recognition accuracy on the test set is less than 90%, a loop optimization mechanism is triggered until the model output is stable and accurate, avoiding feature combination failures caused by overfitting or underfitting.

[0120] Preferably, the prediction unit performs the following steps:

[0121] Step S51: Input the target comprehensive feature vector into the preset analysis model;

[0122] Step S52: Automatically calculate and generate the target function for anthocyanin content using a preset analysis model;

[0123] Step S53: Verify the actual detection results of the samples and adjust the objective function parameters;

[0124] Step S54: Calculate the anthocyanin concentration value of each pixel according to the objective function, and generate the corresponding distribution heatmap data based on the color intensity;

[0125] Step S55: Integrate the objective function and heatmap data into an output result file.

[0126] In this embodiment, a mathematical model is automatically generated by using a pre-defined analysis model driven by a comprehensive feature vector. The weight coefficients are dynamically adjusted through a validation set, and finally, anthocyanin concentration is calculated with pixel-level precision and a heatmap is generated. This achieves non-destructive testing, spatial distribution visualization, and strong robustness in complex environments.

[0127] Preferably, step S52 specifically includes:

[0128] Step S521: Use the target comprehensive feature vector as input data;

[0129] Step S522: Automatically identify the mathematical relationship between the input data and anthocyanin content using a clustering algorithm;

[0130] Step S523: Iteratively optimize the model parameters using known sample data to minimize the error between the predicted result and the actual value;

[0131] Step S524: Output the final formula for calculating anthocyanin content;

[0132] Step S525: Encapsulate the calculation formula into a directly callable function module for subsequent batch calculations or real-time prediction.

[0133] In this embodiment, a clustering algorithm is used to automatically mine the coupling relationship between target features and anthocyanin content, specifically the relationship between color depth and reflection band. The model parameters are then iteratively optimized using known samples to ultimately generate a highly interpretable mathematical formula.

[0134]

[0135] in, The anthocyanin content was calculated using a clustering algorithm. This represents the total number of groups after clustering. This is the cluster group index, with values ​​ranging from 1 to... ; For the first The number of samples within each cluster; For the first The minimum value of the input data for each cluster group; For the first The maximum value of the input data for each cluster group; For continuous input variables; For the first The mean of the input data for each cluster group; For the first The standard deviation of the input data for each cluster group; It is an exponential function, used to construct a weight function in the form of a Gaussian distribution; For the first The sample mean of anthocyanin content is known for each cluster group; For the first The first cluster group Input data The value range is 1 to .

[0136] Calculated by formula The value represents the anthocyanin content, and its range is [value range missing]. .when When the correlation strength between the input data of all cluster groups and the mean of the anthocyanin content samples is zero, it means that the samples do not contain anthocyanins; when When the value is higher, it indicates a higher anthocyanin content. The numerator is weighted by Gaussian integral to calculate the data distribution within the cluster, and the denominator is normalized by standard deviation to calculate the data dispersion within the cluster. Overall, it reflects the comprehensive mapping relationship between the input data distribution and the anthocyanin content, which is suitable for rapid quantitative analysis.

[0137] By combining cluster-driven feature-content correlation modeling, dynamic parameter optimization, and formula encapsulation, this approach addresses the issues of poor adaptability and low computational efficiency for complex visible light, spectral, and texture data.

[0138] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An image processing system for identifying the anthocyanin content of colored potatoes, characterized in that, include: Image acquisition unit, illumination correction unit, feature selection unit, and prediction unit; The image acquisition unit is used to acquire visible light images, skin texture images, and spectral images of the potato, and to perform weighted analysis on the images respectively to generate target visible light images, target skin texture images, and target spectral images; the unit includes: The image weighting module is used to stitch together the potato visible light image, skin texture image and spectral image into a feature matrix, and to assign weights according to the information contribution rate of each mode of the feature matrix; The data processing module is used to remove low-weight and high-weight values ​​of the information contribution rate of each mode of the feature matrix, re-integrate the remaining weight values, and establish a new stitching relationship with the potato visible light image, skin texture image and spectral image to construct a weighted covariance matrix. The target output module is used to generate the corresponding target visible light image, target skin texture image and target spectral image from the potato visible light image, skin texture image and spectral image based on the weighted covariance matrix; The illumination correction unit is used to adaptively correct the compensation strategies corresponding to the visible light image, skin texture image and spectral image of the potato according to the light intensity of the surrounding environment of the potato. The feature selection unit is used to fuse the target visible light image, the target epidermal texture image, and the target spectral image to generate target image data, and select the target comprehensive feature vector of the target image data based on the target neural network; The prediction unit performs mathematical calculations based on the target comprehensive feature vector to output the target content function of potato anthocyanins and distribution heatmap data.

2. The image processing system for identifying anthocyanin content in colored potatoes according to claim 1, characterized in that, The image acquisition unit further includes: The image acquisition module uses a visible light camera, a 3D scanner, and a near-infrared spectral imager to acquire visible light images, skin texture images, and spectral images of the potato.

3. The image processing system for identifying anthocyanin content in colored potatoes according to claim 2, characterized in that, The image weighting module performs the following steps: Step S11: Extract the color histogram, gray-level co-occurrence matrix features, and band reflectance curves corresponding to the visible light image, skin texture image, and spectral image of the potato, respectively. Step S12: Input the color histogram, gray-level co-occurrence matrix features and band reflectance curve into a multi-scale feature pyramid network for modal feature interaction to generate a hybrid feature vector containing local details and global semantics. Step S13: Calculate the feature importance score of each modality for the target detection task using a decision tree model, and generate weight coefficients by combining the sensitivity of spectral data to anthocyanin detection. Step S14: Input the weight coefficients into the gating attention mechanism, and perform nonlinear weighting on the feature channels of different modalities through learnable attention weights.

4. The image processing system for identifying anthocyanin content in colored potatoes according to claim 2, characterized in that, The data processing module performs the following steps: Step S21: Standardize all feature data and filter out data that has little impact on the results or fluctuates abnormally. Step S22: Extract color, texture and spectral features respectively, and unify the color, texture and spectral feature information to the same frequency dimension through multi-scale decomposition; Step S23: Combine the multi-source data of the same frequency dimension, calculate the correlation between the multi-source data, and assign different weights according to importance to establish data relationships.

5. The image processing system for identifying anthocyanin content in colored potatoes according to claim 1, characterized in that, The illumination correction unit performs the following steps: Step S31: Simultaneously use a regular camera and a spectrometer to measure the light intensity of the environment around the potato in real time; Step S32: Dynamically adjust the brightness of the visible light image, and extract the fine texture features of the surface and the reflection information of different color bands to initially lock in the anthocyanin features. Step S33: Using intelligent algorithms, the anthocyanin features are amplified and noise in the image is reduced in low light or high contrast conditions. Step S34: Based on the sensitivity of anthocyanins to specific color bands, automatically adjust the weight ratio of each band and jointly optimize the weight ratio.

6. The image processing system for identifying anthocyanin content in colored potatoes according to claim 5, characterized in that, Step S33 specifically includes: Step 331: Separate the illumination component and anthocyanin reflectance component of the low-light image; Step 332: Use a loss function to constrain the structural similarity between the illumination component and the anthocyanin reflection component, generate a pre-processed image, and simultaneously remove random noise from the image; Step 333: Based on the LLCNN architecture, perform multi-channel analysis on the processed image to generate an enhanced image that highlights anthocyanin features; Step 334: Enhance the anthocyanin-related features in the enhanced image and filter out the residual noise again.

7. The image processing system for identifying anthocyanin content in colored potatoes according to claim 1, characterized in that, The feature selection unit performs the following steps: Step S41: Adjust the target visible light image, target skin texture image and target spectral image to a uniform clarity and color range, eliminate the influence of light, and extract the common features of the target visible light image, target skin texture image and target spectral image; Step S42: Analyze the color changes in visible light, extract the surface texture features, and calculate the band data corresponding to the color depth in the spectral image. Step S43: Assign importance based on the correlation degree of target features in each of the target visible light image, target skin texture image and target spectral image, and remove duplicate or irrelevant information; Step S44: The key information extracted from the target visible light image, target skin texture image and target spectral image is spliced ​​into a complete data package, retaining the core data that best represents the characteristics of the target; Step S45: Input the core data into the target neural network model, adjust the parameters of the target neural network model to find the best feature combination, and verify whether the feature combination accurately judges different target features; Step S46: Output the target features and generate a comprehensive feature vector.

8. The image processing system for identifying anthocyanin content in colored potatoes according to claim 7, characterized in that, Step S45 specifically includes: Step S451: Input the core data into the preset model; Step S452: By automatically adjusting the internal weights and biases of the preset model, try different combinations of the features; Step S453: Use the validation set to monitor the discriminative power of the feature combination in real time and record the optimal parameter configuration; Step S454: Use a test set to check whether the preset model can accurately identify different target features; Step S455: If the recognition accuracy does not meet the standard, return to step S452 to readjust the parameters until the feature combination achieves the expected effect.

9. The image processing system for identifying anthocyanin content in colored potatoes according to claim 1, characterized in that, The prediction unit performs the following steps: Step S51: Input the target comprehensive feature vector into the preset analysis model; Step S52: Automatically calculate and generate the objective function for anthocyanin content using the preset analysis model; Step S53: Verify the actual detection results of the sample and adjust the parameters of the objective function; Step S54: Calculate the anthocyanin concentration value of each pixel according to the objective function, and generate corresponding distribution heatmap data based on color intensity; Step S55: Integrate the objective function and heatmap data into an output result file.

10. The image processing system for identifying anthocyanin content in colored potatoes according to claim 9, characterized in that, Step S52 specifically includes: Step S521: Use the target integrated feature vector as input data; Step S522: Automatically identify the mathematical relationship between the input data and anthocyanin content using a clustering algorithm; Step S523: Iteratively optimize the model parameters using known sample data to minimize the error between the predicted result and the actual value; Step S524: Output the final formula for calculating anthocyanin content; Step S525: Encapsulate the calculation formula into a directly callable function module for subsequent batch calculations or real-time predictions.

Images