Field maturity recognition method for tobacco leaf images based on deep semi-supervised transfer learning
Through the deep semi-supervised transfer learning method, using a small number of labeled samples and a large number of unlabeled samples for training, combined with cross entropy and maturity structure constraint loss, the problem of high cost of tobacco leaf maturity recognition in the field is solved, and efficient and accurate recognition effects are achieved.
Patent Information
- Application Number
- CN202310666848.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-06-07
- Publication Date
- 2025-09-19
- Estimated Expiration
- 2043-06-07
AI Technical Summary
In the existing technology, the supervised learning scheme for identifying the maturity of tobacco leaves in the field is difficult to apply in practice because it requires huge manpower and time costs, and image recognition involves multiple scenarios, resulting in excessively high costs.
A method based on deep semi-supervised transfer learning was adopted to establish a tobacco leaf field maturity recognition model by dividing tobacco leaf image samples and using a small number of labeled samples and a large number of unlabeled samples for training. The cross entropy loss, local maximum mean difference loss and maturity structure constraint loss were combined.
It achieves low-cost and efficient identification of tobacco leaf maturity in the field, improves identification accuracy, reduces labeling costs, and is suitable for actual production.
Smart Images
Figure CN116740564B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a method for identifying the maturity of tobacco leaf images in the field, and in particular to a method for identifying the maturity of tobacco leaf images in the field based on deep semi-supervised transfer learning. Background Art
[0002] The maturity of tobacco leaves in the field is a key factor in determining their harvest. Therefore, identifying this maturity has long been a key concern for the tobacco industry. For years, researchers have explored using image sensing technology and machine learning algorithms to address this issue, with supervised learning (SL) algorithms being widely adopted. However, image recognition of tobacco leaf maturity in the field involves a variety of scenarios, including different varieties, planting areas, and planting patterns. Furthermore, SL approaches require significant manpower and time to collect and label large numbers of tobacco images for each scenario. Consequently, these approaches are difficult to implement in practice due to their high financial and time costs. Summary of the Invention
[0003] The purpose of the present invention is to provide a tobacco leaf image field maturity recognition method based on deep semi-supervised transfer learning, which can establish an image recognition model of tobacco leaf field maturity with extremely low annotation cost, and realize low-cost and high-efficiency tobacco leaf image field maturity recognition.
[0004] The present invention adopts the following technical solutions:
[0005] A method for identifying field maturity of tobacco leaf images based on deep semi-supervised transfer learning includes the following steps:
[0006] A: Divide tobacco leaf image samples used for field maturity detection; select some tobacco leaf image samples from each category for labeling. The labeled tobacco leaf image samples are called labeled tobacco leaf image samples in the target domain, and the remaining unlabeled tobacco leaf image samples are called unlabeled tobacco leaf image samples in the target domain. The tobacco leaf image samples are divided into different categories according to the different maturity levels of the tobacco leaves in the field.
[0007] B: A deep backbone network for tobacco leaf field maturity image recognition is established and defined as the target domain network, followed by transfer learning and classifier parameter initialization;
[0008] C: Establish the loss function L of the deep backbone network obtained in step B SSL ;
[0009]
[0010] Among them, L CE is the cross entropy loss, L LMMD is the local maximum mean difference loss, is the tobacco leaf maturity structure constraint loss for labeled tobacco leaf image samples, is the tobacco leaf maturity structure constraint loss of the unlabeled tobacco leaf image samples, λ1, λ2 and λ3 are the weight parameters of the local maximum mean difference loss, the tobacco leaf maturity structure constraint loss of the labeled tobacco leaf image samples and the tobacco leaf maturity structure constraint loss of the unlabeled tobacco leaf image samples, respectively, and f t (·) represents the target domain network, x t,l Indicates that there are labeled tobacco leaf image samples in the target domain, y t,l Indicates the category label of the labeled tobacco leaf image sample in the target domain, represents the probability distribution of labeled tobacco leaf image samples in the target domain, represents the probability distribution of unlabeled tobacco leaf image samples in the target domain, x t,u Represents the unlabeled tobacco leaf image samples in the target domain;
[0011] D: Using the loss function established in step C, the deep backbone network whose parameters have been initialized in step B is trained to obtain the trained deep backbone network.
[0012] E: Using the trained deep backbone network obtained in step D, field maturity recognition is performed on tobacco leaf images to be tested for field maturity.
[0013] In step B, the ResNet50 network is selected as the target domain network.
[0014] In the step B, when performing transfer learning on the deep feature extractor of the deep backbone network, the established deep backbone network is first trained based on a non-tobacco leaf image sample dataset; the selected image dataset is called the source domain, and the deep backbone network obtained after training is called the source domain network; then the feature extractor parameters of the source domain network are migrated to the feature extractor of the target domain network as the initial parameters of the feature extractor of the target domain network.
[0015] In the step B, when initializing the parameters of the Softmax classifier of the deep backbone network, the parameters of the Softmax classifier are initialized according to the following formula:
[0016]
[0017]
[0018] In formula (1), is the initial parameter vector of the Softmax classifier in the target domain network, with the subscript h t Represents the Softmax classifier of the target domain network, and the superscript C represents the initial parameters of the Cth class;
[0019] In formula (2), represent The cth component of function g t (·) represents the feature extractor of the target domain network, It represents that there are labeled tobacco leaf image samples in the target domain, i is the sample indicator variable, represents the i-th labeled tobacco leaf image sample belonging to category c in the target domain, and ||·|| represents the 2-norm operation.
[0020] In step C, the cross entropy loss L CE for:
[0021] in, is the true label of the i-th labeled tobacco leaf image sample, is the predicted label of the i-th labeled tobacco leaf image sample, and K is the number of labeled tobacco leaf image samples.
[0022] In step C, the local maximum mean difference loss L LMMD for:
[0023]
[0024] Where C represents the number of categories of tobacco leaf maturity in the field; M c is the number of labeled samples in category C, is the weight of the i-th labeled sample belonging to the C-th class, k(·) is the kernel function, g t (·) is the feature extractor of the target domain network, There are labeled tobacco leaf image samples in the target domain, i is the sample indicator variable, N c is the number of unlabeled samples predicted to be class C; is the weight of the j-th unlabeled sample predicted to be class c, is an unlabeled tobacco leaf image sample in the target domain, j is a sample indicator variable; the entropy gate function φ(·) is: x * represents the deep embedding vector of the input to the entropy gate function, is the Softmax classifier of the target domain network for x * The predicted probability vector, for The information entropy of , ε is the information entropy threshold; is the c-th dimension component of the true label of the i-th labeled tobacco leaf image sample; is the c-th dimension component of the predicted label of the j-th unlabeled tobacco leaf image sample.
[0025] In the step C, the tobacco leaf maturity structure constraint loss of the labeled tobacco leaf image sample is for:
[0026]
[0027] in, and denote the mean of the depth embedding vectors of unripe, ripe, and overripe labeled tobacco leaf image samples, respectively; cos(·,·) is the cosine distance.
[0028] In the step C, the tobacco leaf maturity structure constraint loss of the unlabeled tobacco leaf image sample for:
[0029]
[0030] in, and They represent the mean of the deep embedding vectors of the unlabeled tobacco leaf image samples predicted as unripe, ripe, and overripe after passing through the entropy gate function.
[0031] In the step D, there is a labeled tobacco leaf image sample x t,l and unlabeled tobacco leaf image sample x t,u The deep backbone network is sent to the first round of training. At this time, the network loss function only uses the cross entropy loss L CE and the local maximum mean difference loss L LMMD After the first round of training, the deep embedding vectors of the current various labeled samples are obtained, and then the similarity of the deep embedding vectors of immature and mature labeled samples is calculated. Similarity of deep embedding vectors of mature and over-mature labeled samples Similarity of deep embedding vectors of mature and over-mature labeled samples Similarity of deep embedding vectors of unlabeled samples predicted as immature and mature Similarity of deep embedding vectors of unlabeled samples predicted as ripe and overripe Similarity to the deep embedding vectors of unlabeled samples predicted as ripe and overripe Then continue to train the deep backbone network; after the i-th round of training, i>1, if and Dissatisfied and Then, the tobacco leaf maturity structure constraint loss of the labeled tobacco leaf image samples is added to the loss function of the i+1th round of training like and Dissatisfied and Then, the tobacco leaf maturity structure constraint loss of the unlabeled tobacco leaf image samples is added to the loss function of the i+1th round of training Until the training is completed, the trained deep backbone network is finally obtained.
[0032] In step B, the maturity of tobacco leaves in the field is divided into three categories: unripe, mature and overripe, that is, C=3 at this time, and the category mark of unripe samples is recorded as 1, the category mark of mature samples is recorded as 2, and the category mark of overripe samples is recorded as 3.
[0033] The present invention proposes a tobacco leaf image field maturity recognition method based on deep semi-supervised transfer learning. Only a very small number of tobacco leaf field maturity labeled images are needed to establish a tobacco leaf field maturity image recognition model, which can be conveniently and flexibly applied to the actual production of tobacco leaf field maturity detection.
[0034] To leverage knowledge from large-scale image datasets, this paper introduces transfer learning techniques into a semi-supervised learning approach, using knowledge learned from these datasets to initialize the semi-supervised learning feature extractor. Furthermore, to mitigate the generalization risk of semi-supervised learning, this paper innovatively proposes a maturity structure constraint. This incorporates innovative prior knowledge of tobacco leaf maturity into the learning process of the recognition algorithm, forcing it to converge to a confidence space and effectively improving the accuracy of identifying tobacco leaf maturity in the field. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 It is a schematic diagram of the process of the present invention. DETAILED DESCRIPTION
[0036] The present invention is described in detail below with reference to the accompanying drawings and embodiments:
[0037] like Figure 1 As shown, the tobacco leaf image field maturity recognition method based on deep semi-supervised transfer learning of the present invention comprises the following steps in sequence:
[0038] A: Division of tobacco leaf image samples used for field maturity detection;
[0039] In this embodiment, the tobacco leaf field maturity image data is defined as the target domain. From the tobacco leaf image samples of each category in the target domain, some tobacco leaf image samples are selected for labeling, and the labeled tobacco leaf image samples are called labeled tobacco leaf image samples x. t,l , the remaining unlabeled tobacco leaf image samples are called unlabeled tobacco leaf image samples x t,u Among them, the categories are divided into three categories according to the maturity of tobacco leaves in the field: unripe, mature and overripe.
[0040] B: Select a deep backbone network for tobacco leaf field maturity image recognition, perform transfer learning on the deep feature extractor of the deep backbone network, and then initialize the parameters of the deep backbone network classifier;
[0041] In the present invention, the deep backbone network established for extracting tobacco field maturity image features is called the target domain network. The target domain network can select a variety of applicable deep image networks, such as VGG network, Inception network and ResNet network; in this embodiment, the ResNet50 network is selected as the target domain network.
[0042] In step B, transfer learning is performed on the deep feature extractor of the established deep backbone network according to the following method, and then the parameters of the Softmax classifier of the deep backbone network are initialized.
[0043] When performing transfer learning on the deep feature extractor of the deep backbone network, the deep backbone network is first trained based on a non-tobacco leaf image sample dataset (such as the ImageNet dataset or other large non-tobacco leaf image datasets); in this embodiment, the selected image dataset is referred to as the source domain, and the deep backbone network obtained after training is referred to as the source domain network; then the feature extractor parameters of the source domain network are migrated to the feature extractor of the target domain network as the initial parameters of the feature extractor of the target domain network.
[0044] When initializing the parameters of the Softmax classifier of the deep backbone network, the parameters of the Softmax classifier are initialized according to the following formula:
[0045]
[0046]
[0047] In formula (1), is the initial parameter vector of the Softmax classifier in the target domain ResNet50 network, with the subscript h t : represents the Softmax classifier of the target domain network. The superscript C represents the initial parameter of the Cth category, that is, the number of categories of tobacco leaf field maturity. In this embodiment, the tobacco leaf field maturity is divided into three categories: unripe, mature, and overripe, that is, C = 3 at this time, and the category label of the unripe sample is recorded as 1, the category label of the mature sample is recorded as 2, and the category label of the overripe sample is recorded as 3.
[0048] In formula (2), represent The cth component of function g t (·) represents the feature extractor of the target domain ResNet50 network, It represents that there are labeled tobacco leaf image samples in the target domain, i is the sample indicator variable, represents the i-th labeled tobacco leaf image sample belonging to category c in the target domain, and ||·|| represents the 2-norm operation.
[0049] C: Establish the loss function L of the deep backbone network obtained in step B SSL ;
[0050]
[0051] Among them, L CE is the cross entropy loss, L LMMD is the local maximum mean difference loss, is the tobacco leaf maturity structure constraint loss for labeled tobacco leaf image samples, is the tobacco leaf maturity structure constraint loss of the unlabeled tobacco leaf image samples, λ1, λ2 and λ3 are the weight parameters of the local maximum mean difference loss, the tobacco leaf maturity structure constraint loss of the labeled tobacco leaf image samples and the tobacco leaf maturity structure constraint loss of the unlabeled tobacco leaf image samples, respectively, and f t (·) represents the target domain network, x t,l Indicates that there are labeled tobacco leaf image samples in the target domain, y t,l Indicates the category label of the labeled tobacco leaf image sample in the target domain, represents the probability distribution of labeled tobacco leaf image samples in the target domain, represents the probability distribution of unlabeled tobacco leaf image samples in the target domain, x t,u Represents the unlabeled tobacco leaf image samples in the target domain;
[0052] The cross entropy loss measures the difference between the true label and the predicted label of the sample. In this embodiment, the cross entropy loss L CE for:
[0053]
[0054] in, is the true label of the i-th labeled tobacco leaf image sample, is the predicted label of the i-th labeled tobacco leaf image sample, and are all one-hot vectors, and K is the number of labeled tobacco leaf image samples;
[0055] In the present invention, the local maximum mean difference loss L LMMD The purpose is to learn from a small number of labeled samples and a large number of unlabeled samples to get a data distribution closer to the real data, making the network model more generalizable. LMMDThe MMD distance between the corresponding categories of samples is calculated, which makes the boundaries of samples of different categories more separable. In this embodiment, the local maximum mean difference loss L LMMD for:
[0056]
[0057] Where C represents the number of categories of tobacco leaf maturity in the field; M c is the number of labeled samples in category C, is the weight of the i-th labeled sample belonging to the C-th class, k(·) is the kernel function, and in this embodiment, the Gaussian kernel function is selected, φ(·) is the entropy gate function, and g t (·) is the feature extractor of the target domain ResNet50 network, There are labeled tobacco leaf image samples in the target domain, i is the sample indicator variable, N c is the number of unlabeled samples predicted to be class C; is the weight of the j-th unlabeled sample predicted to be class c, is an unlabeled tobacco leaf image sample in the target domain, j is a sample indicator variable; the entropy gate function φ(·) determines the confidence of the sample by calculating the entropy of the predicted label, and then selects the sample with higher confidence to participate in the optimization of the objective function. In this embodiment, the entropy gate function φ(·) is: x * represents the deep embedding vector of the input to the entropy gate function, The Softmax classifier of the target domain ResNet50 network is x * The predicted probability vector, for The information entropy of , ε is the information entropy threshold; is the c-th dimension component of the true label of the i-th labeled tobacco leaf image sample; is the c-th dimension component of the predicted label of the j-th unlabeled tobacco leaf image sample;
[0058] In SSL (semi-supervised learning), since the network only uses a small number of labeled samples for learning, these samples are easily overfitted by the network, resulting in poor model generalization. Therefore, the present invention intervenes in the SSL learning process to make it more robust and ultimately build a trustworthy model.
[0059] Since the ripening process of tobacco leaves is actually the aging process of tobacco leaves, there is an increasing relationship between the aging degree in the three maturity categories of unripe, mature and overripe. Therefore, the present invention uses constraints based on the similarity between maturity categories to force the SSL algorithm to converge to the confidence parameter space. The constraints used include two parts: the tobacco leaf maturity structure constraint loss with labeled tobacco leaf image samples and the and tobacco leaf maturity structure constraint loss for unlabeled tobacco leaf image samples
[0060] In this embodiment, there is a tobacco leaf maturity structure constraint loss for the labeled tobacco leaf image samples. for:
[0061]
[0062] in, and denotes the mean of the depth embedding vectors of unripe, ripe, and overripe labeled tobacco leaf image samples respectively; cos(·,·) is the cosine distance, and the cosine distance of two vectors a and b with dimension n is defined as a i and b i Represent the i-th dimension component of vectors a and b respectively;
[0063] In this embodiment, the tobacco leaf maturity structure constraint loss of the unlabeled tobacco leaf image sample for:
[0064]
[0065] in, and They represent the mean of the depth embedding vectors of the unlabeled tobacco leaf image samples predicted as unripe, ripe, and overripe after the entropy gate function;
[0066] D: Using the loss function established in step C, the deep backbone network whose parameters have been initialized in step B is trained to obtain the trained deep backbone network.
[0067] In this embodiment, there are labeled tobacco leaf image samples x t,l and unlabeled tobacco leaf image sample x t,u The deep backbone network is sent to the first round of training. At this time, the network loss function only uses the cross entropy loss L CE and the local maximum mean difference loss L LMMD After the first round of training, the deep embedding vectors of the current various labeled samples are obtained, and then the similarity of the deep embedding vectors of immature and mature labeled samples is calculated. Similarity of deep embedding vectors of mature and over-mature labeled samples Similarity of deep embedding vectors of mature and over-mature labeled samples Similarity of deep embedding vectors of unlabeled samples predicted as immature and mature Similarity of deep embedding vectors of unlabeled samples predicted as ripe and overripe Similarity to the deep embedding vectors of unlabeled samples predicted as ripe and overripe Then continue to train the deep backbone network; after the i-th round of training, i>1, if and Dissatisfied and Then, the tobacco leaf maturity structure constraint loss of the labeled tobacco leaf image samples is added to the loss function of the i+1th round of training like and Dissatisfied and Then, the tobacco leaf maturity structure constraint loss of the unlabeled tobacco leaf image samples is added to the loss function of the i+1th round of training Until the training is completed, the trained deep backbone network is finally obtained;
[0068] E: Using the trained deep backbone network obtained in step D, field maturity recognition is performed on tobacco leaf images to be tested for field maturity.
Claims
1. A method for identifying field maturity of tobacco leaf images based on deep semi-supervised transfer learning, characterized in that: The following steps are included in sequence: A: Divide tobacco leaf image samples used for field maturity detection; select some tobacco leaf image samples from each category for labeling. The labeled tobacco leaf image samples are called labeled tobacco leaf image samples in the target domain, and the remaining unlabeled tobacco leaf image samples are called unlabeled tobacco leaf image samples in the target domain. The tobacco leaf image samples are divided into different categories according to the different maturity levels of the tobacco leaves in the field. B: A deep backbone network for tobacco leaf field maturity image recognition is established and defined as the target domain network, followed by transfer learning and classifier parameter initialization; C: Establish the loss function L of the deep backbone network obtained in step B SSL ; Among them, L CE is the cross entropy loss, L LMMD is the local maximum mean difference loss, is the tobacco leaf maturity structure constraint loss for labeled tobacco leaf image samples, is the tobacco leaf maturity structure constraint loss of the unlabeled tobacco leaf image samples, λ1, λ2 and λ3 are the weight parameters of the local maximum mean difference loss, the tobacco leaf maturity structure constraint loss of the labeled tobacco leaf image samples and the tobacco leaf maturity structure constraint loss of the unlabeled tobacco leaf image samples, respectively, and f t (·) represents the target domain network, x t,l Indicates that there are labeled tobacco leaf image samples in the target domain, y t,l Indicates the category label of the labeled tobacco leaf image sample in the target domain, represents the probability distribution of labeled tobacco leaf image samples in the target domain, represents the probability distribution of unlabeled tobacco leaf image samples in the target domain, x t,u Represents the unlabeled tobacco leaf image samples in the target domain; D: Using the loss function established in step C, the deep backbone network whose parameters have been initialized in step B is trained to obtain the trained deep backbone network. E: Using the trained deep backbone network obtained in step D, field maturity recognition is performed on tobacco leaf images to be tested for field maturity.
2. The method for identifying field maturity of tobacco leaf images based on deep semi-supervised transfer learning according to claim 1, characterized in that: In step B, the ResNet50 network is selected as the target domain network.
3. The method for identifying field maturity of tobacco leaf images based on deep semi-supervised transfer learning according to claim 1, characterized in that: In the step B, when performing transfer learning on the deep feature extractor of the deep backbone network, the established deep backbone network is first trained based on a non-tobacco leaf image sample dataset; the selected image dataset is called the source domain, and the deep backbone network obtained after training is called the source domain network; then the feature extractor parameters of the source domain network are migrated to the feature extractor of the target domain network as the initial parameters of the feature extractor of the target domain network.
4. The method for identifying field maturity of tobacco leaf images based on deep semi-supervised transfer learning according to claim 1, characterized in that: In the step B, when initializing the parameters of the Softmax classifier of the deep backbone network, the parameters of the Softmax classifier are initialized according to the following formula: In formula (1), is the initial parameter vector of the Softmax classifier in the target domain network, with the subscript h t Represents the Softmax classifier of the target domain network, and the superscript C represents the initial parameters of the Cth class; In formula (2), represent The cth component of function g t (·) represents the feature extractor of the target domain network, It represents that there are labeled tobacco leaf image samples in the target domain, i is the sample indicator variable, represents the i-th labeled tobacco leaf image sample belonging to category c in the target domain, and ||·|| represents the 2-norm operation.
5. The method for identifying field maturity of tobacco leaf images based on deep semi-supervised transfer learning according to claim 1, characterized in that: In step C, the cross entropy loss L CE for: in, is the true label of the i-th labeled tobacco leaf image sample, is the predicted label of the i-th labeled tobacco leaf image sample, and K is the number of labeled tobacco leaf image samples.
6. The method for identifying field maturity of tobacco leaf images based on deep semi-supervised transfer learning according to claim 1, characterized in that: In step C, the local maximum mean difference loss L LMMD for: Where C represents the number of categories of tobacco leaf maturity in the field; M c is the number of labeled samples in category C, is the weight of the i-th labeled sample belonging to the C-th class, k(·) is the kernel function, g t (·) is the feature extractor of the target domain network, There are labeled tobacco leaf image samples in the target domain, i is the sample indicator variable, N c is the number of unlabeled samples predicted to be class C; is the weight of the j-th unlabeled sample predicted to be class c, is an unlabeled tobacco leaf image sample in the target domain, j is a sample indicator variable; the entropy gate function φ(·) is: x * represents the deep embedding vector of the input to the entropy gate function, is the Softmax classifier of the target domain network for x * The predicted probability vector, for The information entropy of , ε is the information entropy threshold; is the c-th dimension component of the true label of the i-th labeled tobacco leaf image sample; is the c-th dimension component of the predicted label of the j-th unlabeled tobacco leaf image sample.
7. The method for identifying field maturity of tobacco leaf images based on deep semi-supervised transfer learning according to claim 1, characterized in that: In the step C, the tobacco leaf maturity structure constraint loss of the labeled tobacco leaf image sample is for: in, and denote the mean of the depth embedding vectors of unripe, ripe, and overripe labeled tobacco leaf image samples, respectively; cos(·,·) is the cosine distance.
8. The method for identifying field maturity of tobacco leaf images based on deep semi-supervised transfer learning according to claim 1, characterized in that: In the step C, the tobacco leaf maturity structure constraint loss of the unlabeled tobacco leaf image sample for: in, and They represent the mean of the deep embedding vectors of the unlabeled tobacco leaf image samples predicted as unripe, ripe, and overripe after passing through the entropy gate function.
9. The method for identifying field maturity of tobacco leaf images based on deep semi-supervised transfer learning according to claim 1, characterized in that: In the step D, there is a labeled tobacco leaf image sample x t,l and unlabeled tobacco leaf image sample x t,u The deep backbone network is sent to the first round of training. At this time, the network loss function only uses the cross entropy loss L CE and the local maximum mean difference loss L LMMD After the first round of training, the deep embedding vectors of the current various labeled samples are obtained, and then the similarity of the deep embedding vectors of immature and mature labeled samples is calculated. Similarity of deep embedding vectors of mature and over-mature labeled samples Similarity of deep embedding vectors of mature and over-mature labeled samples Similarity of deep embedding vectors predicted as immature and mature unlabeled samples Similarity of deep embedding vectors of unlabeled samples predicted as ripe and overripe Similarity to the deep embedding vectors of unlabeled samples predicted as ripe and overripe Then continue to train the deep backbone network; after the i-th round of training, i>1, if and Dissatisfied and Then, the tobacco leaf maturity structure constraint loss of the labeled tobacco leaf image samples is added to the loss function of the i+1th round of training like and Dissatisfied and Then, the tobacco leaf maturity structure constraint loss of the unlabeled tobacco leaf image samples is added to the loss function of the i+1th round of training Until the training is completed, the trained deep backbone network is finally obtained.
10. The method for identifying field maturity of tobacco leaf images based on deep semi-supervised transfer learning according to claim 4, characterized in that: In step B, the maturity of tobacco leaves in the field is divided into three categories: unripe, mature and overripe, that is, C=3 at this time, and the category mark of unripe samples is recorded as 1, the category mark of mature samples is recorded as 2, and the category mark of overripe samples is recorded as 3.