A method, system and medium for identifying leaf phenotypes of wild jasmine plants

Abstract

The embodiment of the present application relates to a method, a recognition system and a medium for identifying the leaf phenotype of a wild jasmine plant, and belongs to the technical field of plant intelligent recognition. The recognition method comprises: based on the leaf image of the plant to be tested, using a mixed feature extraction neural network to extract the mixed features of the leaf; based on the mixed features of the leaf, using a discrimination neural network to extract single attribute features, the single attribute features including leaf size features, leaf color features and leaf edge features; based on the mixed features of the leaf and each single attribute feature, using a gating model to determine the fusion parameters corresponding to each single attribute feature; based on each single attribute feature and the fusion parameter, obtaining a fused multi-attribute feature; based on the multi-attribute feature, using an anomaly detection model to perform anomaly detection to obtain the leaf phenotype of the plant to be tested, which can generate a multi-attribute fusion feature with high robustness and complete information attributes, thereby improving the accuracy of identifying abnormal phenotypes of wild jasmine leaves.

Description

Technical Field

[0001] The present application relates to the technical field of plant intelligent identification, and more specifically, to an identification method, an identification system and a medium for identifying the leaf phenotype of a jasmine plant. Background Art

[0002] Wild jasmine plants are mainly distributed in the warm temperate and subtropical regions of the Northern Hemisphere. There are 9 genera and more than 60 species in China, mainly distributed in various places south of the Yangtze River. Wild jasmine plants have many economic values ​​in daily production and life, such as ornamental value (often used in gardening and landscaping), medicinal value (often used in herbal medicine, with anti-inflammatory and detoxification effects), spice extraction (often used in cosmetics and food), nectar plants (can promote the development of beekeeping) and cultural value (in some areas, wild jasmine is closely related to local culture and traditional customs, and has certain cultural and social value). However, the cultivation and planting process of wild jasmine plants is affected by factors such as pests and diseases. Common diseases include powdery mildew and leaf spot. Most pests and diseases usually cause wild jasmine plants to exhibit abnormal phenotypes. With the rapid development of agricultural technology and computer technology, more and more automatic intelligent recognition methods are used to identify abnormal plant phenotypes, but most of them adopt transfer learning methods, and the recognition accuracy and robustness are relatively poor. It is impossible to obtain accurate classification results of diseases and pests of wild jasmine, and it cannot provide important reference for analyzing and controlling wild jasmine plants with abnormal leaf phenotypes. Summary of the invention

[0003] The present application is provided to solve the above defects in the prior art. A method, system and medium for identifying leaf phenotypes of wild jasmine plants are needed, which can generate multi-attribute fusion features with high robustness and complete information attributes, and improve the accuracy of identifying abnormal phenotypes of wild jasmine leaves.

[0004] In a first aspect of the present application, a method for identifying the leaf phenotype of a jasmine plant is provided, and the identification method comprises the following steps. Based on the leaf image of the plant to be tested, a mixed feature extraction neural network is used to extract the mixed features of the leaf. Based on the mixed features of the leaf, a single attribute feature is extracted using an identification neural network, and the single attribute feature includes a leaf size feature, a leaf color feature, and a leaf edge feature. Based on the mixed features of the leaf and each single attribute feature, a gating model is used to determine the fusion parameters corresponding to each single attribute feature. Based on each single attribute feature and the fusion parameter, a fused multi-attribute feature is obtained. Based on the multi-attribute feature, anomaly detection is performed using an anomaly detection model to obtain the leaf phenotype of the plant to be tested.

[0005] In a second aspect of the present application, a system for identifying leaf phenotypes of wild jasmine plants is provided, the identification system comprising a mixed feature extraction module, a multi-attribute feature extraction module, a multi-attribute feature fusion module and a leaf abnormality detection module, the mixed feature extraction module being configured as follows: based on a leaf image of a plant to be tested, a mixed feature of the leaf is extracted using a mixed feature extraction neural network; the multi-attribute feature extraction module being configured as follows: based on the mixed feature of the leaf, a single attribute feature is extracted using a discrimination neural network, the single attribute feature comprising a leaf size feature, a leaf color feature and a leaf edge feature; the multi-attribute feature fusion module being configured as follows: based on the mixed feature of the leaf and each single attribute feature, a gating model is used to determine the fusion parameters corresponding to each single attribute feature; based on each single attribute feature and the fusion parameter, a fused multi-attribute feature is obtained; the leaf abnormality detection module being configured as follows: based on the multi-attribute feature, an anomaly detection model is used to perform anomaly detection to obtain the leaf phenotype of the plant to be tested.

[0006] In a third aspect of the present application, a non-transitory computer-readable medium is provided, on which instructions are stored. When executed by a processor, the instructions execute the steps of the method for identifying the leaf phenotype of a wild jasmine plant as described in any embodiment of the present application.

[0007] The embodiments of the present application provide a method, system and medium for identifying the leaf phenotype of wild jasmine plants. Single attribute features are extracted based on the mixed features of the leaves. The single attribute features include leaf size features, leaf color features and leaf edge features. In this way, multiple attributes of wild jasmine leaves can be obtained. Fusion parameters corresponding to each single attribute feature are determined based on the mixed features of the leaves and each single attribute feature. Fused multi-attribute features are obtained based on each single attribute feature and the fusion parameters. Fine-grained multi-attribute features can be fused through the fusion process to generate multi-attribute mixed features with higher robustness and better information completeness. The abnormal phenotype identification of wild jasmine leaves achieved in this way can more accurately classify abnormal leaf phenotypes, thereby providing an important reference for analyzing and controlling wild jasmine plants with abnormal leaf phenotypes. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In the drawings, which are not necessarily drawn to scale, the same reference numerals may describe similar parts in different views. The same reference numerals with letter suffixes or different letter suffixes may represent different instances of similar parts. The accompanying drawings generally illustrate various embodiments by way of example and not limitation, and together with the specification and claims, are used to illustrate the embodiments applied for. When appropriate, the same reference numerals are used throughout the drawings to refer to the same or similar parts. Such embodiments are illustrative and are not intended to be exhaustive or exclusive embodiments of the present apparatus or method.

[0009] Figure 1 A flow chart showing a method for identifying leaf phenotypes of wild jasmine plants according to an embodiment of the present application;

[0010] Figure 2 A diagram showing the identification process of the leaf phenotype of the wild jasmine plant according to an embodiment of the present application;

[0011] Figure 3 A physical picture showing the leaves of the wild jasmine plant according to an embodiment of the present application;

[0012] Figure 4 A multi-attribute feature visualization diagram of an embodiment of the present application is shown;

[0013] Figure 5 A schematic diagram showing the structure of a system for identifying leaf phenotypes of wild jasmine plants according to an embodiment of the present application. DETAILED DESCRIPTION

[0014] In order to enable those skilled in the art to better understand the technical solution of the present application, the present application is described in detail below in conjunction with the accompanying drawings and specific implementation examples. The embodiments of the present application are further described in detail below in conjunction with the accompanying drawings and specific implementation examples, but are not intended to limit the present application.

[0015] The words "first", "second" and similar words used in this application do not indicate any order, quantity or importance, but are only used to distinguish. The words "include" or "comprise" and similar words mean that the elements before the word include the elements listed after the word, and do not exclude the possibility of including other elements.

[0016] According to an embodiment of the present application, a method for identifying the leaf phenotype of a plant of the genus Jasminum is provided. Figure 1 The flowchart of the method for identifying the leaf phenotype of the wild jasmine plant according to the embodiment of the present application is shown. The identification method comprises the following steps. Figure 2 The following is a diagram showing the identification process of the leaf phenotype of the wild jasmine plant according to an embodiment of the present application. In step 101, based on the leaf image 201 of the plant to be tested, a hybrid feature extraction neural network 202 is used to extract the hybrid feature 203 of the leaf. The leaf image 201 may include multiple images to better extract more feature information about the leaf.

[0017] In step 102, based on the mixed features 203 of the leaves, a single attribute feature is extracted using a discriminant neural network 204, and the single attribute features include a leaf size feature 2051, a leaf color feature 2052, and a leaf edge feature 2053. The leaf size feature 2051 can reflect whether the leaf size is normal. Normal leaf shapes include elliptical or lanceolate shapes, and abnormalities may be manifested as deformities, distortions, or irregular shapes. The leaf color feature 2052 can determine whether the leaf color is normal. Normal leaf colors are dark green, and abnormal colors may be yellowing, brown spots, or mottled linear colors. The leaf edge feature 2053 can determine whether the leaf edge is normal. Normal leaf edges are relatively flat, and abnormal colors may be burnt edges, curled, or damaged. Through the multi-attribute features of the leaves, the abnormal phenotypes of the leaves can be more accurately classified, such as which disease or pest they belong to. Based on each single attribute feature, a multi-attribute fusion feature with high robustness and complete information attributes can be generated.

[0018] In step 103, based on the mixed feature 203 and each single attribute feature of the leaf, the gating model 206 is used to determine the fusion parameters corresponding to each single attribute feature. The mixed feature 203 and the single attribute feature of the leaf can better determine the fusion parameters of the single attribute features relative to the mixed feature, the fusion parameter 2071 corresponding to the leaf size feature 2051, the fusion parameter 2072 corresponding to the leaf color feature 2052, and the fusion parameter 2073 corresponding to the leaf edge feature 2053, thereby controlling the proportion of each single attribute feature included in the fusion feature.

[0019] In step 104, based on the single attribute features and the fusion parameters, a fused multi-attribute feature 208 is obtained. The multi-attribute feature 208 includes information of each single attribute feature, which facilitates abnormality detection.

[0020] In step 105, based on the multi-attribute feature 208, anomaly detection is performed using an anomaly detection model 209 to obtain the leaf phenotype 210 of the plant to be tested. By fusing the multi-attribute feature information of the leaf size feature 2051, the leaf color feature 2052, and the leaf edge feature 2053, it is possible to comprehensively judge whether the leaf phenotype is a normal phenotype or an abnormal phenotype. If it is an abnormal phenotype, the classification of the abnormal leaf phenotype can also be obtained, such as powdery mildew, leaf spot disease, insect pests, etc., thereby providing an important reference for analyzing and controlling wild jasmine plants with abnormal leaf phenotypes. The fusion parameters are determined by the gating model 206 to control the proportion of each single attribute feature to improve the accuracy of intelligent identification of leaf phenotypes.

[0021] In some embodiments, the hybrid feature extraction neural network includes a deep convolutional neural network and a fully connected layer, and extracts the hybrid features of the leaves based on the leaf image of the plant to be tested, including: based on the leaf image of the plant to be tested, using a deep convolutional neural network to extract initial hybrid features; based on the initial hybrid features, using a fully connected layer to perform linear transformation and activation function processing to obtain hybrid features after changing the dimension.

[0022] The processing of the fully connected layer can improve the feature dimension of the mixed features, making the mixed features richer and providing a better analysis basis for attribute feature extraction and anomaly detection.

[0023] In some embodiments, the deep convolutional neural network includes ResNet50, etc. In some embodiments, the formula for fully connected layer processing can be expressed as: , where r i represents the mixed features after linear transformation, σ represents the activation function during linear transformation, represents the weight parameter in the linear transformation process, represents the bias parameter in the linear transformation process, p i represents the mixed features of the leaf image of the plant to be tested, i represents the batch number in the leaf image dataset of the plant to be tested, . is the leaf image dataset of the plant to be tested. The leaf image samples included in the batch. The parameters are optimized through the training process. The activation function σ includes the sigmoid activation function, etc.

[0024] The mixed feature r is obtained by linear transformation i It can be expressed as r i , where bs is used to represent the batch size and 1024 is used to represent the mixed feature dimension after linear transformation.

[0025] The present application includes three identification neural networks, which are respectively used to extract corresponding leaf size features, leaf color features and leaf edge features. That is, one identification neural network corresponds to extracting leaf size features, one identification neural network corresponds to extracting leaf color features, and one identification neural network corresponds to extracting leaf edge features.

[0026] In some embodiments, the recognition method further includes training the discriminative neural network corresponding to each single attribute, and the training method for training the initial discriminative neural network corresponding to each single attribute includes: based on the prediction results of the single attributes of each initial discriminative neural network and the fine-grained attribute annotation content of the leaf image in the leaf image training data set, using the discriminative loss function to train each discriminative neural network to obtain each trained discriminative neural network. In some embodiments, the discriminative neural network may include a fully connected layer.

[0027] The prediction results of the initial identification neural network for each single attribute include: extracting the mixed features after the dimension change to obtain leaf size prediction features, leaf color prediction features and leaf edge prediction features.

[0028] Based on the discrimination loss function, the single attribute features are alienated so that the trained discrimination neural network can obtain the leaf size attribute information, leaf color attribute information and leaf edge attribute information more accurately.

[0029] In some embodiments, the discrimination loss function is:

[0030] ;

[0031] in, represents the fitted data distribution of the leaf image training dataset; represents the actual data distribution of the leaf image training data set; E represents the mathematical expectation; , i represents the leaf image batch number of the leaf image training data set of the plant to be tested; The leaf image training dataset The leaf image samples included in the batch; , represents the prediction results corresponding to the leaf image samples of the i-th batch; It represents the fine-grained attribute annotation content corresponding to the leaf image samples in the i-th batch, and j=1, 2, and 3 represent the annotation content of leaf size, leaf color, and leaf edge, respectively.

[0032] After the identification neural network is trained using the identification loss function, the leaf size characteristics, leaf color characteristics and leaf edge characteristics obtained by the identification neural network are more accurate.

[0033] In some embodiments, based on the mixed features and each single attribute feature of the leaf, a gating model is used to determine the fusion parameters corresponding to each single attribute feature, including: splicing based on the mixed features and each single attribute feature, and performing linear transformation to obtain the gating threshold corresponding to each single attribute feature. In this way, the gating threshold of the mixed feature relative to the single attribute feature can be obtained, thereby controlling the proportion of each single attribute feature.

[0034] In some embodiments, the calculation formula of the gating model is: , where Z i (j) represents the gate threshold; σ represents the activation function in the linear transformation process, W (f) represents the weight parameter in the linear transformation process, b (f) represents the bias parameter in the linear transformation process, r i Represents mixed features; r i (j) Represents a single attribute feature, j=1,2,3, respectively representing different single attribute features, , i represents the leaf image batch number of the leaf image dataset of the plant to be tested. The activation function σ can be a Tanh (hyperbolic tangent) function, etc.

[0035] In some embodiments, obtaining the fused multi-attribute features based on the individual single-attribute features and the fusion parameters includes the following calculation process:

[0036] ;

[0037] , where r i (j) Represents a single attribute feature; Z i (j) represents the gating threshold; h i (j) represents the fused single attribute features; j=1,2,3, respectively representing different single attribute features; h i represents the fused multi-attribute features, , i represents the leaf image batch number of the leaf image dataset of the plant to be tested.

[0038] The leaf phenotype includes whether it is normal. If it is abnormal, it may be due to disease or insect pest. Different abnormal situations correspond to different single attribute characteristics. The proportion of each single attribute is controlled by the gating threshold to obtain a more accurate leaf phenotype result.

[0039] like Figure 3 This is a real picture of the leaves of wild jasmine plants. Figure 4 It is a visualization diagram of the fused multi-attribute features. Figure 3and Figure 4 The multi-attribute features that can be obtained include the attribute features of leaf size, leaf color, and leaf edge.

[0040] In some embodiments, the anomaly detection model may include a fully connected layer. The classification result of the leaf phenotype is obtained by the anomaly detection model.

[0041] In some embodiments, the recognition method further comprises: using the classification loss function to extract the initial hybrid feature neural network (the parameters involved are represented by θ (h) ), the initial discriminant neural network (the parameters involved are represented by θ (m) ), the initial gating model (the parameters involved are represented by θ (f) ), the initial anomaly detection model (the parameters involved are represented by θ (d) ) to obtain the trained hybrid feature extraction neural network, identification neural network, gating model, and anomaly detection model. Accurate parameter values ​​are obtained through training.

[0042] In some embodiments, the training method includes: determining a leaf image training data set and a test set of a plant to be tested, wherein the leaf image training data set includes normal leaf images and abnormal leaf images; annotating the abnormal leaf images; and training a deep convolutional neural network and a fully connected layer using the annotated abnormal leaf images and normal leaf images. The parameters of the deep convolutional neural network and the fully connected layer are obtained through training. Here, the content of the data annotation of the abnormal leaf images is the leaf phenotype annotation.

[0043] In some embodiments, data annotation of abnormal leaf images includes: fine-grained attribute annotation and leaf phenotype annotation of leaf images in the leaf image training data set. The annotation content of the fine-grained attribute annotation includes whether the leaf shape is abnormal, whether the leaf color is abnormal, and whether the leaf edge is abnormal. The annotated leaf phenotype includes normal, powdery mildew, leaf spot and / or insect pests, etc. For example, p k represents the leaf image training dataset, , 1≤k≤N, where R is used to represent a real number set, 3 is used to represent the number of channels, and 224×224 is used to represent the leaf image size. The leaf images with fine-grained attribute annotations are used to train the fine-grained attribute extraction of the identification neural network. When annotating the leaf images in the leaf image training dataset, each leaf image can be annotated with fine-grained attributes and leaf phenotypes, or the abnormal leaves in the first dataset can be annotated with fine-grained attributes, and the abnormal leaves in the second dataset can be annotated with leaf phenotypes. The first dataset and the second dataset can be independent or overlapping.

[0044] In some embodiments, the leaf image training data set may include 200 normal leaf images and 600 abnormal leaf images. The number of leaf images may be variable.

[0045] In some embodiments, the classification loss function is:

[0046] , where E represents the mathematical expectation; y i Indicates the leaf phenotype annotation content; represents the leaf phenotype prediction result, , i represents the leaf image batch number of the leaf image training data set of the plant to be tested. Based on the classification loss function, the initial hybrid feature extraction neural network (the parameters involved are expressed as θ(h)), the initial discrimination neural network (the parameters involved are expressed as θ(m)), the initial gating model (the parameters involved are expressed as θ(f)), and the initial anomaly detection model (the parameters involved are expressed as θ(d)) are trained to obtain the parameters corresponding to the hybrid feature extraction neural network, the discrimination neural network, the gating model, and the anomaly detection model. The classification loss function is minimized, and the optimized parameters are obtained, which can be expressed as follows.

[0047]

[0048] The prediction result is the prediction result of the leaf phenotype obtained by the abnormal detection module for the leaf image training data set, including normal, powdery mildew, leaf spot and / or insect pests, etc. The output of the real leaf phenotype of the same plant to be tested may be one or more. The leaf phenotype annotation content is the real leaf phenotype annotated on the leaf image of the leaf image training data set of the plant to be tested, including normal, powdery mildew, leaf spot and / or insect pests, etc. The real leaf phenotype of the same plant to be tested may be one or more.

[0049] Based on the real labeling results and prediction results, the initial hybrid feature extraction neural network, the initial discrimination neural network, the initial gating model, and the initial anomaly detection model are trained, so that the output results of the trained hybrid feature extraction neural network, the discrimination neural network, the gating model, and the anomaly detection model are closer to the actual leaf phenotype.

[0050] According to an embodiment of the present application, a system for identifying leaf phenotypes of wild jasmine plants is also provided. Figure 5As shown, the recognition system 500 includes a mixed feature extraction module 501, a multi-attribute feature extraction module 502, a multi-attribute feature fusion module 503 and a leaf anomaly detection module 504. The mixed feature extraction module 501 is configured as follows: based on the leaf image of the plant to be tested, the mixed features of the leaf are extracted using a mixed feature extraction neural network; the multi-attribute feature extraction module 502 is configured as follows: based on the mixed features of the leaf, single attribute features are extracted using a discriminant neural network, and the single attribute features include leaf size features, leaf color features and leaf edge features; the multi-attribute feature fusion module 503 is configured as follows: based on the mixed features of the leaf and each single attribute feature, the fusion parameters corresponding to each single attribute feature are determined using a gating model; based on each single attribute feature and the fusion parameters, the fused multi-attribute features are obtained; the leaf anomaly detection module 504 is configured as follows: based on the multi-attribute features, anomaly detection is performed using an anomaly detection model to obtain the leaf phenotype of the plant to be tested.

[0051] By fusing the multi-attribute feature information of leaf size, leaf color, and leaf edge, it is possible to comprehensively judge whether the leaf phenotype is normal or abnormal. If it is an abnormal phenotype, it is also possible to classify the abnormal leaf phenotype, such as powdery mildew, leaf spot, insect pests, etc., thus providing an important reference for analyzing and controlling wild jasmine plants with abnormal leaf phenotypes. The fusion parameters are determined by the gating model to control the proportion of each single attribute feature to improve the accuracy of intelligent identification of leaf phenotypes.

[0052] In some embodiments, the recognition system includes a multi-attribute training module, which is configured as follows: based on the prediction results of the single attributes of each initial discriminant neural network and the fine-grained attribute annotation content of the leaf image in the leaf image training data set, each discriminant neural network is trained using a discriminant loss function to obtain each trained discriminant neural network.

[0053] In some embodiments, the recognition system also includes a leaf phenotype training module, and the leaf phenotype training module is configured to: use a classification loss function to train an initial hybrid feature extraction neural network, an initial discrimination neural network, an initial gating model, and an initial anomaly detection model to obtain a trained hybrid feature extraction neural network, a discrimination neural network, a gating model, and anomaly detection model.

[0054] According to an embodiment of the present application, a non-transitory computer-readable medium is also provided, on which instructions are stored. When executed by a processor, the instructions execute the steps of the method for identifying the leaf phenotype of a wild jasmine plant as described in any embodiment of the present application.

[0055] Verification of wild jasmine leaf image data set: The training learning parameters are set as follows: the training batch size is set to 128, the initial learning rate size is set to 10-3, the learning rate decay cycle is set to 0.1, and the Adam optimizer is used for 100 training cycles to obtain the hybrid feature extraction neural network, identification neural network, gating model, and anomaly detection model involved in the method for identifying the leaf phenotype of wild jasmine plants after training. Table 1 shows the experimental results of the technology, where RF is used to represent the random forest algorithm, SVM represents the support vector machine algorithm, DBN represents the deep belief network, LSTM is used to represent the long short-term memory network, CNN is used to represent the Resnet convolutional neural network, and MCGU is used to represent the method for identifying the leaf phenotype of wild jasmine plants composed of the hybrid feature extraction neural network, identification neural network, gating model, and anomaly detection model after the above training of this application.

[0056] Table 1 Experimental results

[0057]

[0058] Therefore, the identification method of the present application has a high accuracy rate and can provide an important basis for determining the types of diseases of wild jasmine plants and the degree of their damage.

[0059] In addition, although exemplary embodiments have been described herein, the scope includes any and all embodiments based on the present application with equivalent elements, modifications, omissions, combinations (e.g., various embodiments intersecting schemes), adaptations or changes. The elements in the claims are to be interpreted broadly based on the language adopted in the claims, and are not limited to the examples described in this specification or during the practice of this application, and the examples are to be interpreted as non-exclusive. Therefore, this specification and examples are intended to be considered as examples only, and the true scope and spirit are indicated by the following claims and the full scope of their equivalents.

[0060] The above description is intended to be illustrative rather than restrictive. For example, the above examples (or one or more schemes thereof) may be used in combination with each other. For example, a person of ordinary skill in the art may use other embodiments when reading the above description. In addition, in the above-mentioned specific embodiments, various features may be grouped together to simplify the present application. This should not be interpreted as an intention that a feature of an application that does not require protection is necessary for any claim. On the contrary, the subject matter of the present application may be less than all the features of an embodiment of a particular application. Thus, the following claims are incorporated herein in the specific embodiments as examples or embodiments, wherein each claim is independently a separate embodiment, and it is considered that these embodiments may be combined with each other in various combinations or arrangements. The scope of the present invention should be determined with reference to the appended claims and the full scope of equivalent forms granted by these claims.

[0061] The above embodiments are only exemplary embodiments of the present application and are not intended to limit the present invention. The protection scope of the present invention is defined by the claims. Those skilled in the art may make various modifications or equivalent substitutions to the present invention within the essence and protection scope of the present application, and such modifications or equivalent substitutions shall also be deemed to fall within the protection scope of the present invention.

Claims

1. A method for identifying leaf phenotypes of wild jasmine plants, characterized in that: The identification method comprises: Based on the leaf image of the plant to be tested, the initial mixed features are extracted using a deep convolutional neural network; based on the initial mixed features, a fully connected layer is used to perform linear transformation and activation function processing to obtain the mixed features after the dimension is changed; Based on the mixed features of the leaves, a single attribute feature is extracted using a discriminant neural network, wherein the single attribute feature includes a leaf size feature, a leaf color feature, and a leaf edge feature; a training method for training the initial discriminant neural network corresponding to each single attribute includes: based on the prediction results of the single attributes of each initial discriminant neural network and the fine-grained attribute annotation content of the leaf image in the leaf image training data set, each discriminant neural network is trained using a discriminant loss function to obtain each discriminant neural network after training; Based on the mixed features and the single attribute features of the leaves, the gating model is used to determine the fusion parameters corresponding to the single attribute features, including: splicing based on the mixed features and the single attribute features, and performing linear transformation to obtain the gating thresholds corresponding to the single attribute features; Obtaining fused multi-attribute features based on the individual single-attribute features and fusion parameters; Based on the multi-attribute features, anomaly detection is performed using an anomaly detection model to obtain the leaf phenotype of the plant to be tested.

2. The identification method according to claim 1, characterized in that: The identification loss function is: ； in, represents the fitted data distribution of the leaf image training dataset; represents the actual data distribution of the leaf image training data set; E represents the mathematical expectation; , i represents the leaf image batch number of the leaf image training data set of the plant to be tested; The leaf image training dataset The leaf image samples included in the batch; , represents the prediction results corresponding to the leaf image samples of the i-th batch; It represents the fine-grained attribute annotation content corresponding to the leaf image samples in the i-th batch, and j=1, 2, and 3 represent the annotation content of leaf size, leaf color, and leaf edge, respectively.

3. The identification method according to claim 1, characterized in that: The identification method also includes: using a classification loss function to train an initial hybrid feature extraction neural network, an initial discrimination neural network, an initial gating model, and an initial anomaly detection model to obtain a trained hybrid feature extraction neural network, a discrimination neural network, a gating model, and anomaly detection model.

4. A system for identifying leaf phenotypes of wild jasmine plants, characterized in that: The recognition system includes a hybrid feature extraction module, a multi-attribute feature extraction module, a multi-attribute training module, a multi-attribute feature fusion module and a leaf abnormality detection module, wherein the hybrid feature extraction module is configured to: extract initial hybrid features based on the leaf image of the plant to be tested using a deep convolutional neural network; Based on the initial mixed features, a fully connected layer is used to perform linear transformation and activation function processing to obtain a mixed feature after the dimension is changed; the multi-attribute feature extraction module is configured as follows: based on the mixed features of the leaves, a discriminant neural network is used to extract single attribute features, and the single attribute features include leaf size features, leaf color features, and leaf edge features; The multi-attribute training module is configured as follows: based on the prediction results of the single attributes of each initial discriminant neural network and the fine-grained attribute annotation content of the leaf image in the leaf image training data set, each discriminant neural network is trained using a discriminant loss function to obtain each discriminant neural network after training; The multi-attribute feature fusion module is configured as follows: based on the mixed features and the single attribute features of the leaves, the gating model is used to determine the fusion parameters corresponding to the single attribute features, including: splicing based on the mixed features and the single attribute features, and performing linear transformation to obtain the gating thresholds corresponding to the single attribute features; and obtaining the fused multi-attribute features based on the single attribute features and the fusion parameters; The leaf anomaly detection module is configured to: perform anomaly detection based on the multi-attribute features using an anomaly detection model to obtain the leaf phenotype of the plant to be tested.

5. The identification system according to claim 4, characterized in that: The recognition system also includes a leaf phenotype training module, which is configured to: use a classification loss function to train an initial hybrid feature extraction neural network, an initial discrimination neural network, an initial gating model, and an initial anomaly detection model to obtain a trained hybrid feature extraction neural network, a discrimination neural network, a gating model, and anomaly detection model.

6. A non-transitory computer-readable medium having instructions stored thereon, which, when executed by a processor, execute the steps of the method for identifying the leaf phenotype of a plant of the genus Jasminum as claimed in any one of claims 1 to 3.

Images