Wild pear variety identification system and method based on leaf image deep learning

Abstract

The present application provides a wild pear variety identification system and method based on leaf image deep learning, which relates to the technical field of deep learning algorithm, including obtaining a leaf enhanced image, mixing an original leaf image training set with a generated image after format conversion, standardizing the mixed image for preprocessing, inputting it to a constructed multi-scale residual network for training, and outputting a wild pear variety probability vector. Obtain a training set, test set and validation set of wild pear leaf images under different light conditions, input the original leaf training set into the constructed generative adversarial network to obtain a leaf enhanced image, increase the data volume of the sample, perform two-dimensional discrete wavelet transform on the mixed image to improve the recognition accuracy of the leaf image, input it to the constructed multi-scale residual network for training, introduce an attention mechanism, strengthen the key area, obtain an optimized wild pear variety classification model, and output a wild pear variety probability vector.

Description

Technical Field

[0001] The present invention relates to the technical field of deep learning algorithms, and in particular to a wild pear variety identification system and method based on deep learning of leaf images. Background Art

[0002] In modern agriculture and plant science, wild pears, as an important gene pool for fruit trees, are crucial for their conservation and utilization. Wild pears not only contribute significantly to ecological stability but also provide valuable genetic resources for pear breeding, genetic improvement, and stress tolerance research. With growing awareness of biodiversity, the need for identifying and classifying wild pear varieties is increasing. Accurately identifying different wild pear varieties is crucial for effectively managing and protecting these resources and promoting their application in ecological restoration and agricultural production.

[0003] However, existing methods for identifying wild pear varieties primarily rely on traditional morphological feature analysis and manual identification. These methods often require the experience and knowledge of professionals, are cumbersome, inefficient, and susceptible to human influence. For example, the identification of external features such as leaf color, shape, and texture often relies on the observer's subjective judgment, leading to inconsistent identification results. Furthermore, many traditional methods exhibit poor accuracy and robustness when dealing with complex backgrounds and diverse leaf features, failing to meet the requirements for efficient and accurate variety identification.

[0004] To overcome these shortcomings, image processing and deep learning technologies have been gradually introduced into the field of plant classification in recent years. While some studies have utilized image recognition technology for plant classification, most employ simple feature extraction methods, failing to fully exploit the advantages of deep learning models in feature learning. As a result, in practical applications, the recognition accuracy of these models remains insufficient, particularly given the complexity and diversity of leaf images, making it difficult to effectively distinguish between different wild pear varieties. Therefore, developing an efficient recognition method that fully leverages the advantages of deep learning has become a key task in current plant science research.

[0005] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art. Summary of the Invention

[0006] The purpose of the present invention is to provide a wild pear variety identification system and identification method based on leaf image deep learning to solve the problems raised in the above background technology.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] The wild pear variety identification method based on leaf image deep learning includes the following steps:

[0009] S1: Select the multi-scale residual network as the baseline model and build a model for wild pear variety recognition. The model includes an input layer, an intermediate layer, and an output layer. The generative adversarial network is used as the input layer, the image preprocessing module is used as the first intermediate layer, the multi-scale residual network is used as the second intermediate layer, and the output of the model is used as the output layer. The wild pear leaf image dataset is obtained and annotated to construct the training set, test set, and validation set.

[0010] S2: Input the original leaf training set into the model and perform data enhancement using the generative adversarial network constructed by the input layer to obtain leaf enhanced data;

[0011] S3: Use the first-layer image preprocessing module in the middle of the model to convert the generated leaf enhanced image format, mix the original leaf image training set with the generated image after format conversion, and perform standardization preprocessing on the mixed image;

[0012] S4: The preprocessed mixed image is input into the multi-scale residual network constructed in the second middle layer for training. During the training process, the loss function is optimized at the same time.

[0013] S5: The output layer uses the validation set and the test set to evaluate the performance of the wild pear variety classification model and is used to identify the wild pear variety after the model meets the requirements.

[0014] Furthermore, the wild pear leaf image dataset is obtained and annotated to construct a training set, a test set and a validation set. Specifically, the annotation is performed according to the minimum external rectangle of each leaf, and the annotated wild pear leaf image datasets of different types are divided into training sets, test sets and validation sets under different lighting conditions in a ratio of 7:2:1.

[0015] Furthermore, the constructed generative adversarial network includes generator network construction and discriminator design, wherein the generator network adopts a U-Net structure and a PatchGAN discriminator, and the original leaf training set is input into the constructed generative adversarial network to obtain a leaf enhanced image.

[0016] Furthermore, the format conversion of the generated leaf enhanced image includes numerical normalization and size unification, and retaining its RGB three-channel data.

[0017] Furthermore, the standardization preprocessing includes mixing the original leaf image training set with the generated image after format conversion in a 3:1 ratio, performing a two-dimensional discrete wavelet transform on each input image according to the channel, decomposing it into low-frequency subbands and high-frequency subbands, performing adaptive gain amplification on the high-frequency subbands, and performing an inverse wavelet transform on the enhanced high-frequency subbands and the original low-frequency subbands to reconstruct them into a spatial domain image, performing mean zeroing and variance normalization on the RGB channels respectively:

[0018]

[0019] Where I represents the input feature map of the multi-scale residual network layer l; I mix Represents mixed input image; DWT discrete wavelet transform; DWT(I mix ) high represents the high-frequency subband set; H represents the high-frequency subband information entropy; E represents the high-frequency subband energy; λ represents the enhancement intensity coefficient; IDWT represents the inverse wavelet transform; μ represents the global mean of the mixed data set; σ represents the global standard deviation of the mixed data set; ∈ represents the data stability value; I real represents the original leaf image training set; I fake represents the generated image after format conversion; LL represents the low-frequency subband.

[0020] Furthermore, the constructed multi-scale residual network includes designing a three-way parallel asymmetric convolution kernel to capture different leaf morphological features, and introducing a spatial attention mechanism to construct a residual function to transfer the leaf morphological features within the residual block.

[0021]

[0022] in,

[0023]

[0024] y represents the output feature map; Conv2D 1×1 Represents a 1×1 convolutional layer; Conv2D 3×3 Represents a 3×3 convolutional layer; Conv2D 5×3 (x l ) represents a 5×3 convolutional layer; W1 represents the weight of the first fully connected layer; W2 represents the weight of the second fully connected layer; represents the eigenvalue of the cth channel at the spatial position (h, w); C k (x l )(h, w) represents the eigenvalue of the kth branch at channel c and position (h, w); h represents the height coordinate; w represents the width coordinate; H represents the feature map height; W represents the feature map width. represents the fusion feature; F(x l θl ) represents the residual function; Represents the spatial attention weight map.

[0025] Furthermore, the wild pear variety identification method based on leaf image deep learning is characterized in that the wild pear variety probability vector is:

[0026] satisfy

[0027] p k represents the probability that the input leaf belongs to the kth category; K represents the total number of wild pear leaf species.

[0028] Furthermore, the optimization loss function includes optimizing the classification loss function and the attention supervision loss function, where

[0029]

[0030] in,

[0031]

[0032] in, represents the classification loss function; represents the attention supervision loss function; y i represents the true label of the i-th sample; p i Indicates the probability that the model predicts that the i-th sample belongs to the true category; A(x l )(h, w) represents the prediction weight of the l-th layer spatial attention mechanism at position (h, w); M gt (h, w) represents the manually annotated true attention continuous value mask.

[0033] Furthermore, the use of the validation set and the test set to evaluate the performance of the wild pear variety classification model includes, during the training process, using the validation set to perform model evaluation at the end of each training cycle, monitoring the training process of the model, and after the training is completed, using the test set to perform a final evaluation of the model, and counting the number of samples predicted as correct but actually incorrect for each sample, the number of samples predicted as correct but actually incorrect, and the number of samples predicted as incorrect but actually correct, and calculating the overall precision and recall of the test set, wherein the overall precision is as follows:

[0034]

[0035] Among them, TP represents the number of samples that are predicted to be correct and are actually correct; FP represents the number of samples that are predicted to be correct but are actually incorrect;

[0036] The recall rates are as follows:

[0037]

[0038] Among them, Recall represents the recall rate; FN represents the number of samples that are predicted to be incorrect but are actually correct.

[0039] The present invention further provides a wild pear variety identification system based on leaf image deep learning, wherein the system is used to perform the above-mentioned identification method, comprising:

[0040] A model construction module is used to select a multi-scale residual network as a baseline model and build a model for wild pear variety recognition. The model includes an input layer, an intermediate layer, and an output layer. The generative adversarial network is used as the input layer, the image preprocessing module is used as the first intermediate layer, the multi-scale residual network is used as the second intermediate layer, and the output of the model is used as the output layer. A wild pear leaf image dataset is obtained and annotated to construct a training set, a test set, and a validation set.

[0041] The model training module is used to input the original leaf training set into the model and perform data enhancement using the generative adversarial network constructed by the input layer to obtain leaf enhanced data;

[0042] A model processing module is used to convert the format of the generated leaf enhancement image using the first-layer image preprocessing module in the middle of the model, mix the original leaf image training set with the generated image after format conversion, and perform standardization preprocessing on the mixed image;

[0043] The loss optimization module is used to input the preprocessed mixed image into the multi-scale residual network constructed in the second intermediate layer for training. During the training process, the loss function is optimized at the same time.

[0044] Model output module, the output layer uses the validation set and test set to evaluate the performance of the wild pear variety classification model, and is used to identify the wild pear variety after the model meets the requirements.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] The present invention constructs a combination of a multi-scale residual network and a generative adversarial network, and the model can demonstrate stronger robustness and accuracy when processing diverse leaf images. The deep learning model can automatically extract and learn leaf features, avoiding the subjectivity brought by traditional manual feature extraction methods. Through the introduction of standardized preprocessing and spatial attention mechanism, the model can more accurately capture the morphological characteristics of leaves, thereby significantly improving the ability to recognize different varieties of wild pears. The use of a generative adversarial network for data enhancement can significantly increase the diversity and richness of the training set. By generating realistic and diverse leaf images, the model can be effectively trained with a smaller sample size, solving the limitations and cost issues in the data acquisition process. This method not only improves the training effect of the model, but also effectively reduces the risk of overfitting, making the model more robust on the validation set and test set. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of the overall method flow of the present invention;

[0048] Figure 2 Schematic diagram of the overall system structure of the present invention. DETAILED DESCRIPTION

[0049] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to specific embodiments.

[0050] It should be noted that, unless otherwise defined, the technical or scientific terms used in the present invention should have the usual meanings understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative position relationships. When the absolute position of the object being described changes, the relative position relationship may also change accordingly.

[0051] Example:

[0052] See also Figure 1 , the present invention provides a technical solution:

[0053] The wild pear variety identification method based on leaf image deep learning includes the following steps:

[0054] S1: Select the multi-scale residual network as the baseline model and build a model for wild pear variety recognition. The model includes an input layer, an intermediate layer, and an output layer. The generative adversarial network is used as the input layer, the image preprocessing module is used as the first intermediate layer, the multi-scale residual network is used as the second intermediate layer, and the output of the model is used as the output layer. The wild pear leaf image dataset is obtained and annotated to construct the training set, test set, and validation set.

[0055] A wild pear leaf image dataset was obtained and annotated to construct training, test, and validation sets. Specifically, the minimum outer rectangle of each leaf was used for annotation, and the annotated wild pear leaf image datasets of different types were divided into training, test, and validation sets under different lighting conditions in a ratio of 7:2:1.

[0056] S2: Input the original leaf training set into the model and perform data enhancement using the generative adversarial network constructed by the input layer to obtain leaf enhanced data;

[0057] The constructed generative adversarial network includes generator network construction and discriminator design, wherein the generator network adopts U-Net structure and PatchGAN discriminator, and the original leaf training set is input into the constructed generative adversarial network to obtain the leaf enhanced image.

[0058] S3: Use the first-layer image preprocessing module in the middle of the model to convert the generated leaf enhanced image format, mix the original leaf image training set with the generated image after format conversion, and perform standardization preprocessing on the mixed image;

[0059] Standardization preprocessing involves mixing the original leaf image training set with the generated image after format conversion in a 3:1 ratio. For each input image, a two-dimensional discrete wavelet transform is performed on each channel to decompose it into low-frequency subbands and high-frequency subbands. The high-frequency subbands are adaptively amplified, and the enhanced high-frequency subbands and the original low-frequency subbands are inversely transformed by wavelet to reconstruct the spatial domain image. The three RGB channels are subjected to mean zeroing and variance normalization respectively:

[0060]

[0061] Where I represents the input feature map of the multi-scale residual network layer l; I mix Represents mixed input image; DWT discrete wavelet transform; DWT(I mix ) high represents the high-frequency subband set; H represents the high-frequency subband information entropy; E represents the high-frequency subband energy; λ represents the enhancement intensity coefficient; IDWT represents the inverse wavelet transform; μ represents the global mean of the mixed data set; σ represents the global standard deviation of the mixed data set; ∈ represents the data stability value; I realrepresents the original leaf image training set; I fake represents the generated image after format conversion; LL represents the low-frequency subband.

[0062] Among them, for the real image I real and generate image I fake Perform wavelet transform to separate low-frequency sub-band LL and high-frequency sub-band; dynamically calculate the enhancement coefficient based on the information entropy H and energy E of the high-frequency sub-band Strengthen complex texture areas; reconstruct the image by inverse LL transforming the enhanced high frequency and the original low frequency; and perform mean μ and normalization σ on the output to improve the stability of model training.

[0063] S4: The preprocessed mixed image is input into the multi-scale residual network constructed in the second middle layer for training. During the training process, the loss function is optimized at the same time.

[0064] The constructed multi-scale residual network includes designing a three-way parallel asymmetric convolution kernel to capture the morphological features of different leaves, introducing a spatial attention mechanism, constructing a residual function, and transferring the morphological features of the leaves within the residual block.

[0065]

[0066]

[0067] in,

[0068]

[0069] y represents the output feature map; Conv2D 1×1 Represents a 1×1 convolutional layer; Conv2D 3×3 Represents a 3×3 convolutional layer; Conv2D 5×3 (x l ) represents a 5×3 convolutional layer; W1 represents the weight of the first fully connected layer; W2 represents the weight of the second fully connected layer; represents the eigenvalue of the cth channel at the spatial position (h, w); C k (x l )(h, w) represents the eigenvalue of the kth branch at channel c and position (h, w); h represents the height coordinate; w represents the width coordinate; H represents the feature map height; W represents the feature map width. represents the fusion feature; F(x l θ l ) represents the residual function; represents the spatial attention weight.

[0070] The spatial attention mechanism includes the channel attention path and the spatial attention path. The channel weight is obtained by global pooling and then transformed into a fully connected layer to obtain the channel attention weight:

[0071]

[0072] The number of channels is compressed from 256 to 128 through feature transformation, and spatial weights are generated:

[0073]

[0074] The channel attention weight s c With the spatial attention map A(x l ) combined with:

[0075]

[0076] Add the weighted features to the original input to get the next layer of features:

[0077]

[0078] The probability vector of wild pear varieties is:

[0079]

[0080] p k represents the probability that the input leaf belongs to the kth category; K represents the total number of wild pear leaf species.

[0081] Optimizing the loss function includes optimizing the classification loss function and the attention supervision loss function, where

[0082]

[0083] in,

[0084]

[0085] in, represents the classification loss function; represents the attention supervision loss function; y i represents the true label of the i-th sample; p i Indicates the probability that the model predicts that the i-th sample belongs to the true category; A(x l )(h, w) represents the prediction weight of the l-th layer spatial attention mechanism at position (h, w); M gt (h, w) represents the manually annotated true attention continuous value mask.

[0086] S5: The output layer uses the validation set and the test set to evaluate the performance of the wild pear variety classification model and is used to identify the wild pear variety after the model meets the requirements.

[0087] The performance of the wild pear variety classification model is evaluated using the validation set and the test set, including using the validation set to evaluate the model at the end of each training cycle during the training process, monitoring the training process of the model, using the test set to perform a final evaluation of the model after the training is completed, and counting the number of samples predicted as correct samples, the number of samples predicted as correct but actually incorrect, and the number of samples predicted as incorrect but actually correct, calculating the overall precision and recall of the test set, wherein the overall precision is as follows:

[0088]

[0089] Wherein TP represents the number of samples predicted correctly and actually correctly; FP represents the number of samples predicted correctly but actually incorrectly;

[0090] The true label of the sample is compared with the label predicted by the model, and the overall accuracy is calculated, the higher the accuracy, the more obvious the recognition effect of the model, and the lower the accuracy, the worse the recognition effect of the model;

[0091] The recall is as follows:

[0092]

[0093] Wherein Recall represents the recall; FN represents the number of samples predicted as incorrect but actually correct.

[0094] The greater the recall, the more obvious the classification effect of the model, and the smaller the recall, the worse the classification effect of the model.

[0095] Please refer to Figure 2 The application further provides a wild pear variety identification system based on leaf image deep learning, which is used to execute the identification method described above, and comprises:

[0096] A model construction module is configured to select a multi-scale residual network as a baseline model, construct a model for wild pear variety identification, and include an input layer, an intermediate layer and an output layer in the model, use a generative adversarial network as the input layer, use a picture preprocessing module as the first layer of the intermediate layer, use a multi-scale residual network as the second layer of the intermediate layer, use the output of the model as the output layer, and obtain a wild pear leaf image data set to label and construct a training set, a test set and a validation set.

[0097] A model training module is configured to input an original leaf training set into the model, use the generative adversarial network constructed by the input layer to perform data enhancement, and obtain leaf enhanced data.

[0098] A model processing module is used to convert the format of the generated leaf enhancement image using the first-layer image preprocessing module in the middle of the model, mix the original leaf image training set with the generated image after format conversion, and perform standardization preprocessing on the mixed image;

[0099] The loss optimization module is used to input the preprocessed mixed image into the multi-scale residual network constructed in the second intermediate layer for training. During the training process, the loss function is optimized at the same time.

[0100] Model output module, the output layer uses the validation set and test set to evaluate the performance of the wild pear variety classification model, and is used to identify the wild pear variety after the model meets the requirements.

[0101] The above formulas are all dimensionless and numerical calculations. The formulas are obtained by collecting a large amount of data and performing software simulation to obtain the most recent real situation. The preset parameters in the formulas are set by technicians in this field according to actual conditions.

[0102] The above embodiments can be implemented in whole or in part by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented in whole or in part in the form of a computer program product. Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software depends on the specific application and design constraints of the technical solution.

[0103] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, and may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment as needed.

[0104] The above is only a specific implementation method of the present application, but the scope of protection of the present application is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed in this application, which should be covered by the scope of protection of the present application.

Claims

1. A wild pear variety identification method based on leaf image deep learning, characterized in that: include: S1: Select the multi-scale residual network as the baseline model and build a model for wild pear variety recognition. The model includes an input layer, an intermediate layer, and an output layer. The generative adversarial network is used as the input layer, the image preprocessing module is used as the first intermediate layer, the multi-scale residual network is used as the second intermediate layer, and the output of the model is used as the output layer. The wild pear leaf image dataset is obtained and annotated to construct the training set, test set, and validation set. S2: Input the original leaf training set into the model and perform data enhancement using the generative adversarial network constructed by the input layer to obtain leaf enhanced data; S3: Use the first-layer image preprocessing module in the middle of the model to convert the generated leaf enhanced image format, mix the original leaf image training set with the generated image after format conversion, and perform standardization preprocessing on the mixed image; The standardization preprocessing includes mixing the original leaf image training set with the generated image after format conversion in a 3:1 ratio, performing a two-dimensional discrete wavelet transform on each input image according to the channel, decomposing it into low-frequency sub-bands and high-frequency sub-bands, performing adaptive gain amplification on the high-frequency sub-bands, and performing an inverse wavelet transform on the enhanced high-frequency sub-bands and the original low-frequency sub-bands to reconstruct them into a spatial domain image, and performing mean zeroing and variance normalization on the RGB channels respectively: Among them, x l Represents the input feature map of the multi-scale residual network layer l; I mix Represents mixed input image; DWT discrete wavelet transform; DWT(I mix ) high represents the high-frequency subband set; H represents the high-frequency subband information entropy; E represents the high-frequency subband energy; λ represents the enhancement intensity coefficient; IDWT represents the inverse wavelet transform; μ represents the global mean of the mixed data set; σ represents the global standard deviation of the mixed data set; ∈ represents the data stability value; I real represents the original leaf image training set; I fake represents the generated image after format conversion; LL represents the low-frequency subband; S4: The preprocessed mixed image is input into the multi-scale residual network constructed in the second middle layer for training. During the training process, the loss function is optimized at the same time. The constructed multi-scale residual network includes designing a three-way parallel asymmetric convolution kernel to capture the morphological features of different leaves. Among them, a 1×1 convolution layer is used to extract the edge features of the leaf, a 3×3 convolution layer is used to capture the overall shape of the leaf, and a 5×3 convolution layer is used to extract the structure of the entire leaf. In addition, a spatial attention mechanism is introduced to construct a residual function to transfer the morphological features of the leaf, including the shape, size, texture and color of the leaf, within the residual block. in, y represents the output feature map; Conv2D 1×1 Represents a 1×1 convolutional layer; Conv2D 3×3 Represents a 3×3 convolutional layer; Conv2D 5×3 (x l ) represents a 5×3 convolutional layer; W1 represents the weight of the first fully connected layer; W2 represents the weight of the second fully connected layer; represents the eigenvalue of the cth channel at the spatial position (h, w); C k (x l )(h,w) represents the eigenvalue of the k-th branch at channel c and position (h,w); h represents the height coordinate; w represents the width coordinate; H represents the feature map height; W represents the feature map width; represents the fusion feature; F(x l θ l ) represents the residual function; represents the spatial attention weight map; S5: The output layer uses the validation set and the test set to evaluate the performance of the wild pear variety classification model and is used to identify the wild pear variety after the model meets the requirements.

2. The wild pear variety identification method based on leaf image deep learning according to claim 1, characterized in that, The wild pear leaf image dataset is obtained and annotated to construct a training set, a test set and a validation set. Specifically, the annotation is performed according to the minimum external rectangle of each leaf, and the annotated wild pear leaf image datasets of different types are divided into training sets, test sets and validation sets under different lighting conditions in a ratio of 7:2:

1.

3. The wild pear variety identification method based on leaf image deep learning according to claim 1, characterized in that, The constructed generative adversarial network includes generator network construction and discriminator design, wherein the generator network adopts a U-Net structure and a PatchGAN discriminator, and the original leaf training set is input into the constructed generative adversarial network to obtain a leaf enhanced image.

4. The wild pear variety identification method based on leaf image deep learning according to claim 1, characterized in that The format conversion of the generated leaf enhanced image includes numerical normalization and size unification, and retains its RGB three-channel data.

5. The method for identifying wild pear varieties based on leaf image deep learning according to claim 1, characterized in that: The probability vector of wild pear varieties is: p k represents the probability that the input leaf belongs to the kth category; K represents the total number of wild pear leaf species.

6. The method for identifying wild pear varieties based on leaf image deep learning according to claim 1, characterized in that: The optimization loss function includes optimizing the classification loss function and the attention supervision loss function, where in, in, represents the classification loss function; represents the attention supervision loss function; y i represents the true label of the i-th sample; p i Indicates the probability that the model predicts that the i-th sample belongs to the true category; A(x l )(h,w) represents the prediction weight of the l-th layer spatial attention mechanism at position (h,w); M gt (h, w) represents the manually annotated true attention continuous value mask.

7. The method for identifying wild pear varieties based on leaf image deep learning according to claim 1, characterized in that: The method of using the validation set and the test set to evaluate the performance of the wild pear variety classification model includes: performing model evaluation at the end of each training cycle using the validation set during the training process, monitoring the training process of the model, and finally evaluating the model using the test set after the training is completed, and counting the number of samples predicted to be correct but actually incorrect, and the number of samples predicted to be incorrect but actually correct, and calculating the overall precision and recall of the test set, wherein the overall precision is as follows: Among them, TP represents the number of samples that are predicted to be correct and are actually correct; FP represents the number of samples that are predicted to be correct but are actually incorrect; The recall rates are as follows: Among them, Recall represents the recall rate; FN represents the number of samples that are predicted to be incorrect but are actually correct.

8. A wild pear variety identification system based on leaf image deep learning, characterized by: The system is used to perform the identification method according to any one of claims 1 to 7, comprising: A model construction module is used to select a multi-scale residual network as a baseline model and build a model for wild pear variety recognition. The model includes an input layer, an intermediate layer, and an output layer. The generative adversarial network is used as the input layer, the image preprocessing module is used as the first intermediate layer, the multi-scale residual network is used as the second intermediate layer, and the output of the model is used as the output layer. A wild pear leaf image dataset is obtained and annotated to construct a training set, a test set, and a validation set. The model training module is used to input the original leaf training set into the model and perform data enhancement using the generative adversarial network constructed by the input layer to obtain leaf enhanced data; A model processing module is used to convert the format of the generated leaf enhancement image using the first layer preprocessing module in the middle of the model, mix the original leaf image training set with the generated image after format conversion, and perform standardization preprocessing on the mixed image; The loss optimization module is used to input the preprocessed mixed image into the multi-scale residual network constructed in the second intermediate layer for training. During the training process, the loss function is optimized at the same time. Model output module, the output layer uses the validation set and test set to evaluate the performance of the wild pear variety classification model, and is used to identify the wild pear variety after the model meets the requirements.

Images