Lycium barbarum quality screening model training method and system
By building a generative adversarial network, generating virtual comparison data samples and training screening models, the problem of insufficient number of sample images in the machine vision screening model is solved, and the classification accuracy and robustness of wolfberry quality screening is improved.
Patent Information
- Application Number
- CN202510206697.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-25
- Publication Date
- 2025-06-10
AI Technical Summary
When using machine vision screening models, due to the number of sample images collected, sufficient feature image data cannot be obtained, resulting in the quality screening effect of wolfberry that cannot meet production needs.
The target image of wolfberry samples is collected through a machine vision camera, the image parameters and comparison parameters of the recognition feature are extracted, the error parameters are calculated as the input of the generator G, and a generative adversarial network is constructed to generate virtual comparison data samples, and thus generate virtual image samples for training and screening models.
It effectively solves the problem of insufficient number of sample images, obtains sufficient feature image data, improves the classification accuracy and operational robustness of the screening model, and meets the demand for screening effect during production and processing.
Smart Images
Figure CN120125935A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data processing, and particularly relates to a method and system for training a wolfberry quality screening model. Background Art
[0002] The quality of fruits is an important standard for grading, and the quality grading indicators of fruits include external quality and internal quality. Traditional manual and simple mechanical grading methods cannot meet the grading requirements of the internal and external quality of fruits. Machine vision technology has the advantages of large information processing capacity and strong processing ability; moreover, it can effectively avoid artificial mechanical damage, etc., and is more conducive to the quality grading of fruits. Applying machine vision technology to the quality screening of wolfberries can significantly improve the object recognition efficiency. A fruit quality grading system based on machine vision generally consists of an image acquisition module, an image processing and analysis module, and a grading module. Compared with the grading methods of manual and simple machinery, the automation level of machine vision technology is higher. By setting appropriate detection indicators through a program, such as the geometric, color, texture, and law texture energy and other characteristics of the fruit image to be detected, the fruit can be detected without damage, and the detection accuracy is also more accurate.
[0003] Currently, when adopting a machine vision screening model, it is necessary to sample the images of the items to be detected to obtain feature images at multiple scales, so as to improve the classification accuracy and operation robustness of the screening model. In theory, the larger the amount of sampled image data and the more complete the types of images, the better the model training effect and classification effect. However, in the actual process, limited by the number of sampled images collected, it is often impossible to obtain sufficient feature image data for pre-training the model, resulting in the wolfberry quality screening effect not meeting the production requirements. For this reason, we propose a method and system for training a wolfberry quality screening model. Summary of the Invention
[0004] The main purpose of the present invention is to provide a method and system for training a wolfberry quality screening model, which can effectively solve the problems in the background art.
[0005] To achieve the above purpose, the technical solution adopted by the present invention is as follows:
[0006] A method for training a wolfberry quality screening model, comprising:
[0007] Collecting a target image of a wolfberry sample through a machine vision camera of a screening device, and extracting the image parameter x of the recognition feature for wolfberry sample grade screening in the target image ij , where x ij represents the image parameter of the j-th recognition feature of the i-th wolfberry sample, and obtaining the comparison parameter x of the j-th recognition feature of the i-th wolfberry sample ij', where \(i = 1, 2, \ldots, n\); \(n\) is the number of wolfberry samples; \(j = 1, 2, \ldots, m\); \(m\) is the number of recognition feature items; the recognition features include at least one of shape features, color features, texture features, spatial features, local features, and global features;
[0008] According to the image parameter \(x\) ij and the comparison parameter \(x\) ij ', calculate and obtain the error parameter \(x\) of the \(j\)-th recognition feature of the \(i\)-th wolfberry sample ij , where \(x\) ij = x ij - x ij ';
[0009] Using the error parameter \(x\) ij as the input of the generator \(G\), and using the obtained comparison parameter \(x\) ij ' as the real data to construct a generative adversarial network. Use the generator \(G\) to generate virtual comparison data \(G(z)\) of the recognition features. Through the comparison parameter \(x\) ij ' and the generated virtual comparison data \(G(z)\), train the discriminator \(D\) and the generator \(G\) of the generative adversarial network until the output result of the discriminator \(D\) is \(D[G(z)] = 1\). Then use the trained generator \(G\) model to synthesize several virtual comparison data samples of the recognition features;
[0010] According to the obtained virtual comparison data samples, generate virtual image samples with each combination of the recognition features. Use the generated virtual image samples as the training samples for the screening model of the screening device, and train the screening model through the training samples.
[0011] The training process of the generative adversarial network includes the following steps:
[0012] S21: Through the generation network of the generator \(G\), generate virtual comparison data samples according to the input error parameter \(x\) ij , and input them into the discriminator \(D\), and use the discriminator network of the discriminator \(D\) to make a judgment;
[0013] S22: The discriminator network classifies the input virtual comparison data samples and the comparison parameter \(x\) ij ', and calculates the classification error. Among them, the classification error is characterized by any of the following indicators, including:
[0014] Indicator 1: Accuracy rate, the calculation formula is: Accuracy rate = Number of correctly classified samples / Total number of classified samples × 100%;
[0015] Indicator 2: Error rate, calculated as: Error rate = (Total number of classification samples - Number of correctly classified samples) / Total number of classification samples × 100%;
[0016] S23: Update the network parameters of the discriminant network according to the classification error to improve its classification accuracy;
[0017] S24: The generation network updates its own parameters according to the feedback information of the discriminant network to generate more realistic virtual comparison data samples;
[0018] S25: Repeat the above steps S21 - S24 until the preset number of training rounds or convergence conditions are reached.
[0019] The training process of the generator G includes the following steps:
[0020] S31: Obtain the error parameter x ij ;
[0021] S32: Transform the error parameter x ij ;
[0022] S33: Determine the classification result of the discriminator D, and divide the classification result into "true" or "false". Among them, the output of the discriminator D is a binary variable D(x) = 1 or D(x) = 0, and,
[0023] When the output D(x) = 1, it indicates that the classification result of the discriminator D is true;
[0024] When the output D(x) = 0, it indicates that the classification result of the discriminator D is false;
[0025] S34: Calculate the loss of the discriminant network according to the classification result of the discriminator D;
[0026] S35: Through the discriminator D and the generator G for backpropagation to obtain the gradient of the discriminant network;
[0027] S36: Use the gradient of the discriminant network to change the generation network parameters of the generator G.
[0028] The training process of the discriminator D includes the following steps:
[0029] S41: Classify the input comparison parameter x ij ' and the virtual comparison data sample generated by the generator G;
[0030] S42: The loss function of the discriminant network of the discriminator D punishes the mistakes it makes;
[0031] S43: Through backpropagation, the discriminator D updates the weights of its discriminant network; wherein, the discriminator D is used to discriminate the authenticity of the input data, that is, to distinguish between real data and virtual comparison data generated by the generator G; it is assumed that the input parameter of the discriminator D is x, and the output parameter is D(x), and D(x) represents the probability that x is real data.
[0032] When the output parameter D(x) = 1, it means that the input parameter x is 100% real data.
[0033] When the output parameter D(x) = 0, it means that the input parameter x cannot be real data.
[0034] A wolfberry quality screening model training system includes a target image acquisition module, a feature parameter acquisition module, an error parameter acquisition module, a generative adversarial network construction module, and a virtual image generation module.
[0035] The target image acquisition module is used to acquire the target image of the wolfberry sample through the machine vision camera of the screening device.
[0036] The feature parameter acquisition module is used to extract the image parameter x of the recognition feature for wolfberry sample grade screening in the target image ij , and obtain the comparison parameter x ij ' of the j-th recognition feature of the i-th wolfberry sample.
[0037] The error parameter acquisition module is used to calculate and obtain the error parameter x ij of the j-th recognition feature of the i-th wolfberry sample according to the image parameter x ij and the comparison parameter x ij .
[0038] The generative adversarial network construction module is used to use the error parameter x ij as the input of the generator G, and use the obtained comparison parameter x ij ' as real data to construct a generative adversarial network. Among them, the generative adversarial network uses the generator G to generate virtual comparison data G(z) of the recognition feature, and trains the discriminator D and the generator G of the generative adversarial network through the comparison parameter x ij ' and the generated virtual comparison data G(z) until the output result of the discriminator D is D[G(z)] = 1, and then uses the trained generator G model to synthesize several virtual comparison data samples of the recognition feature.
[0039] The virtual image generation module is configured to generate virtual image samples with each combination of the recognition features according to the acquired virtual comparison data samples, use the generated virtual image samples as training samples for the screening model of the screening device, and train the screening model with the training samples.
[0040] The system includes a memory, a processor, and a computer program stored on the memory and executable on the processor.
[0041] The present invention has the following beneficial effects.
[0042] Compared with the prior art, the target image of the wolfberry sample is collected by the machine vision camera of the screening device, and the image parameter x of the recognition feature for the grade screening of the wolfberry sample in the target image is extracted. ij And the comparison parameter x of the j-th recognition feature of the i-th wolfberry sample is obtained. ij ', according to the image parameter x ij and the comparison parameter x ij ', the error parameter x of the j-th recognition feature of the i-th wolfberry sample is calculated and obtained. ij Taking the error parameter x ij as the input of the generator G, using the obtained comparison parameter x ij ' as the real data to construct a generative adversarial network, using the generator G to generate virtual comparison data G(z) of the recognition features, and training the discriminator D and the generator G of the generative adversarial network with the comparison parameter x ij ' and the generated virtual comparison data G(z). Until the output result of the discriminator D is D[G(z)] = 1, use the trained generator G model to synthesize several virtual comparison data samples of the recognition features, generate virtual image samples with each combination of the recognition features according to the virtual comparison data samples, use the generated virtual image samples as training samples for the screening model of the screening device, and train the screening model with the training samples, which can effectively solve the problem of insufficient quantity of the collected sample images in the actual process, thereby obtaining sufficient feature image data for pre-training the model, improving the classification accuracy and running robustness of the screening model, and meeting the requirements for the screening effect in the production and processing process. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1 It is a flowchart of a method for training a wolfberry quality screening model according to the present invention.
[0044] Figure 2 It is a structural block diagram of a wolfberry quality screening model training system according to the present invention.
[0045] Figure 3 It is a structural block diagram of the generative adversarial network constructed in the technical solution of the present invention.
[0046] Figure 4 Schematic diagram of the training process of the generative adversarial network constructed in the technical solution of the present invention;
[0047] Figure 5 Schematic diagram of the training process of the generator G of the generative adversarial network constructed in the present invention;
[0048] Figure 6 Schematic diagram of the training process of the discriminator D of the generative adversarial network constructed in the present invention. Detailed implementation manners
[0049] The following further describes the present invention in conjunction with the detailed implementation manners. Among them, the attached drawings are only for illustrative purposes, showing only schematic diagrams, rather than physical diagrams, and should not be construed as a limitation to the present invention. In order to better illustrate the detailed implementation manners of the present invention, some components in the attached drawings will be omitted, enlarged or reduced, which do not represent the size of the actual product.
[0050] The implementation process of the technical solution of the present invention includes the following steps:
[0051] Step 1: Collect the target image of the goji berry sample through the machine vision camera of the screening device;
[0052] Step 2: Extract the image parameter x of the recognition feature for goji berry sample grade screening in the target image ij , where x ij represents the image parameter of the j-th recognition feature of the i-th goji berry sample, and obtain the comparison parameter x ij ' of the j-th recognition feature of the i-th goji berry sample, where i = 1, 2,..., n; n is the number of goji berry samples; j = 1, 2,..., m; m is the number of items of the recognition feature;
[0053] The recognition feature includes at least one of shape feature, color feature, texture feature, spatial feature, local feature, and global feature;
[0054] Step 3: Calculate and obtain the error parameter x ij of the j-th recognition feature of the i-th goji berry sample according to the image parameter x ij and the comparison parameter x ij , where x ij = x ij - x ij ';
[0055] Step 4: Use the error parameter x ij as the input of the generator G, and use the obtained comparison parameter x ij ' as the real data to construct as Figure 3The shown generative adversarial network uses the generator G of the generative adversarial network to generate virtual comparison data G(z) for recognition features, and through the comparison parameter x ij ' and the generated virtual comparison data G(z), train the discriminator D and the generator G of the generative adversarial network until the output result of the discriminator D is D[G(z)] = 1, and use the trained generator G model to synthesize several virtual comparison data samples of recognition features;
[0056] Among them, as Figure 4 shown, the training process of the generative adversarial network includes the following steps:
[0057] S21: Through the generation network of the generator G, generate virtual comparison data samples according to the input error parameter x ij and input them into the discriminator D, and use the discriminator network of the discriminator D to make a judgment;
[0058] S22: The discriminator network classifies the input virtual comparison data samples and the comparison parameter x ij ', and calculates the classification error. Among them, the classification error is characterized by any of the following indicators, including:
[0059] Indicator 1: Accuracy rate, the calculation formula is: Accuracy rate = Number of correctly classified samples / Total number of classified samples × 100%;
[0060] Indicator 2: Error rate, the calculation formula is: Error rate = (Total number of classified samples - Number of correctly classified samples) / Total number of classified samples × 100%;
[0061] S23: According to the classification error, update the network parameters of the discriminator network to improve its classification accuracy;
[0062] S24: The generation network updates its own parameters according to the feedback information of the discriminator network to generate more realistic virtual comparison data samples;
[0063] S25: Repeat the above steps S21 - S24 until the preset number of training rounds or convergence conditions are reached.
[0064] As Figure 5 shown, the training process of the generator G includes the following steps:
[0065] S31: Obtain the error parameter x ij ;
[0066] S32: Transform the error parameter x ij ;
[0067] S33: Determine the classification result of discriminator D, and divide the classification result into "true" or "false". Among them, the output of discriminator D is a binary variable D(x) = 1 or D(x) = 0, and,
[0068] When the output D(x) = 1, it indicates that the classification result of discriminator D is true;
[0069] When the output D(x) = 0, it indicates that the classification result of discriminator D is false;
[0070] S34: Calculate the loss of the discriminant network according to the classification result of discriminator D;
[0071] S35: Perform backpropagation through discriminator D and generator G to obtain the gradient of the discriminant network;
[0072] S36: Use the gradient of the discriminant network to change the generation network parameters of generator G.
[0073] As Figure 6 shown, the training process of discriminator D includes the following steps:
[0074] S41: Classify the input comparison parameter x ij ' and the virtual comparison data sample generated by generator G;
[0075] S42: The loss function of the discriminant network of discriminator D punishes the mistakes it makes;
[0076] S43: Through backpropagation, discriminator D will update the weights of its discriminant network; among them, discriminator D is used to determine the authenticity of the input data, that is, to distinguish real data and virtual comparison data generated by generator G; it is set that the input parameter of discriminator D is x, and the output parameter is D(x), and D(x) represents the probability that x is real data;
[0077] When the output parameter D(x) = 1, it means that the input parameter x is 100% real data;
[0078] When the output parameter D(x) = 0, it means that the input parameter x cannot be real data.
[0079] Step 5: Generate virtual image samples with various recognition feature combinations according to the obtained virtual comparison data samples;
[0080] Step 6: Use the generated virtual image samples as the training samples of the screening model for the screening device, and train the screening model with the training samples.
[0081] The foregoing has shown and described the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments, and what is described in the above embodiments and the specification is only to illustrate the principle of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.
Claims
1. A wolfberry quality screening model training method, characterized in that: include: The target image of the wolfberry sample is collected by the machine vision camera of the screening device, and the image parameter x in the target image of the identification feature for the grade screening of the wolfberry sample is extracted. ij , where x ij It is expressed as the image parameter of the jth identification feature of the i-th wolfberry sample, and the contrast parameter x of the jth identification feature of the i-th wolfberry sample is obtained. ij ', where i = 1, 2, ..., n; n is the number of wolfberry samples; j = 1, 2, ..., m; m is the number of identification features; According to the image parameter x ij and the comparison parameter x ij 'Calculate the error parameter x of the jth identification feature of the i-th wolfberry sample ij , where x ij =x ij -x ij '; With the error parameter x ij As the input of the generator G, to obtain the contrast parameter x ij 'Use the real data to build a generative adversarial network, use the generator G to generate the virtual comparison data G(z) of the recognition feature, and use the comparison parameter x ij ' and the generated virtual comparison data G(z) to train the discriminator D and generator G of the generative adversarial network until the output result of the discriminator D is D[G(z)]=1, and then use the trained generator G model to synthesize a number of virtual comparison data samples of the recognition features; According to the acquired virtual comparison data samples, virtual image samples with each combination of the identification features are generated, and the generated virtual image samples are used as training samples for the screening model of the screening device, and the screening model is trained by the training samples.
2. A wolfberry quality screening model training method according to claim 1, characterized in that: The training process of the generative adversarial network includes the following steps: S21: The error parameter x is input through the generator network of the generator G. ij Generate virtual comparison data samples and input them into the discriminator D, and use the discriminant network of the discriminator D to make judgments; S22: The discriminant network inputs the virtual comparison data sample and the comparison parameter x ij 'Perform classification and calculate the classification error; S23: updating the network parameters of the discriminant network according to the classification error to improve its classification accuracy; S24: the generating network updates its own parameters according to the feedback information of the discriminating network to generate more realistic virtual comparison data samples; S25: Repeat the above steps S21 to S24 until a preset number of training rounds or convergence condition is reached.
3. A wolfberry quality screening model training method according to claim 1, characterized in that: The training process of the generator G includes the following steps: S31: Obtain the error parameter x ij ; S32: For the error parameter x ij Make a transformation; S33: Determine the classification result of the discriminator D and classify the classification result into "true" or "false", wherein the output of the discriminator D is a binary variable D(x)=1 or D(x)=0, and, When the output D(x) = 1, it indicates that the classification result of the discriminator D is true; When the output D(x) = 0, it indicates that the classification result of the discriminator D is false; S34: Calculate the loss of the discriminant network according to the classification result of the discriminator D; S35: Back propagate through the discriminator D and the generator G to obtain the gradient of the discriminator network; S36: Use the gradient of the discriminator network to change the generation network parameters of the generator G.
4. A wolfberry quality screening model training method according to claim 1, characterized in that: The training process of the discriminator D includes the following steps: S41: Input the comparison parameter x ij 'Classify the virtual comparison data samples generated by the generator G; S42: The loss function of the discriminator network of the discriminator D punishes the mistakes it makes; S43: Through back propagation, the discriminator D updates the weights of its discriminant network; wherein the discriminator D is used to discriminate the authenticity of the input data, that is, to distinguish the real data from the virtual comparison data generated by the generator G; the input parameter of the discriminator D is set to be x, and the output parameter is D(x), where D(x) represents the probability that x is the real data; When the output parameter D(x) = 1, it means that the input parameter x is 100% true data; When the output parameter D(x)=0, it means that the input parameter x cannot be real data.
5. A wolfberry quality screening model training method according to claim 1, characterized in that: The identification features include at least one of shape features, color features, texture features, space features, local features, and global features.
6. A wolfberry quality screening model training method according to claim 2, characterized in that: The classification error is characterized by any of the following indicators, including: Indicator 1: Accuracy, calculated as follows: Accuracy = Number of correctly classified samples / Total number of classified samples × 100%; Indicator 2: Error rate, calculated as follows: Error rate = (total sample size for classification - number of samples correctly classified) / total sample size for classification × 100%.
7. A wolfberry quality screening model training system, characterized in that: It includes a target image acquisition module, a feature parameter acquisition module, an error parameter acquisition module, a generative adversarial network construction module, and a virtual image generation module; The target image acquisition module is used to acquire the target image of the wolfberry sample through the machine vision camera of the screening device; The feature parameter acquisition module is used to extract the image parameter x of the target image in the identification feature for wolfberry sample grade screening. ij , where x ij It is expressed as the image parameter of the jth identification feature of the i-th wolfberry sample, and the contrast parameter x of the jth identification feature of the i-th wolfberry sample is obtained. ij ', where i = 1, 2, ..., n; n is the number of wolfberry samples; j = 1, 2, ..., m; m is the number of identification features; the identification features include at least one of shape features, color features, texture features, spatial features, local features, and global features; The error parameter acquisition module is used to obtain the error parameter according to the image parameter x ij and the comparison parameter x ij 'Calculate the error parameter x of the jth identification feature of the i-th wolfberry sample ij , where x ij =x ij -x ij '; The generative adversarial network building module is used to generate ij As the input of the generator G, to obtain the contrast parameter x ij 'As real data to build a generative adversarial network, wherein the generative adversarial network uses the generator G to generate virtual comparison data G(z) of the recognition feature, and uses the comparison parameter x ij ' and the generated virtual comparison data G(z) to train the discriminator D and generator G of the generative adversarial network until the output result of the discriminator D is D[G(z)]=1, and then use the trained generator G model to synthesize a number of virtual comparison data samples of the recognition features; The training process of the generative adversarial network includes the following steps: The error parameter x is input through the generator network of the generator G. ij Generate virtual comparison data samples and input them into the discriminator D, and use the discriminant network of the discriminator D to make judgments; The discriminant network inputs the virtual comparison data sample and the comparison parameter x ij 'Classify and calculate the classification error, wherein the classification error is characterized by any of the following indicators, including: Indicator 1: Accuracy, calculated as follows: Accuracy = Number of correctly classified samples / Total number of classified samples × 100%; Indicator 2: Error rate, calculated as follows: Error rate = (total sample size for classification - number of samples correctly classified) / total sample size for classification × 100%; According to the classification error, updating the network parameters of the discriminant network to improve its classification accuracy; The generation network updates its own parameters according to the feedback information of the discriminant network to generate more realistic virtual comparison data samples; Repeat the above iterations until the preset number of training rounds or convergence condition is reached; The training process of the generator G includes the following steps: Get the error parameter x ij ; For the error parameter x ij Make a transformation; The classification result of the discriminator D is judged and classified as "true" or "false", wherein the output of the discriminator D is a binary variable D(x)=1 or D(x)=0, and, When the output D(x) = 1, it indicates that the classification result of the discriminator D is true; When the output D(x) = 0, it indicates that the classification result of the discriminator D is false; Calculate the loss of the discriminant network based on the classification results of the discriminator D; Back propagates through the discriminator D and the generator G to obtain the gradient of the discriminator network; Use the gradient of the discriminator network to change the parameters of the generator network G; The training process of the discriminator D includes the following steps: The comparison parameter x for the input ij 'Classify the virtual comparison data samples generated by the generator G; The loss function of the discriminator network of the discriminator D penalizes the mistakes it makes; Through back propagation, the discriminator D will update the weights of its discriminant network; wherein, the discriminator D is used to discriminate the authenticity of the input data, that is, to distinguish the real data from the virtual comparison data generated by the generator G; the input parameter of the discriminator D is x, and the output parameter is D(x), where D(x) represents the probability that x is the real data; When the output parameter D(x) = 1, it means that the input parameter x is 100% true data; When the output parameter D(x) = 0, it means that the input parameter x cannot be real data; The virtual image generation module is used to generate virtual image samples with each combination of the identification features based on the acquired virtual comparison data samples, so as to use the generated virtual image samples as training samples for the screening model of the screening device, and to train the screening model through the training samples.
8. A wolfberry quality screening model training system according to claim 7, characterized in that: The system comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the steps of the method according to any one of claims 1 to 6 are implemented when the processor executes the program.