Hawthorn disease identification method based on deep convolutional neural network and disease image
By using a hawthorn disease identification method based on deep convolutional neural networks, the problem of time-consuming and labor-intensive traditional detection methods has been solved. This method enables high-precision automated identification and early warning of hawthorn diseases, thereby improving production efficiency and fruit quality.
Patent Information
- Application Number
- CN202511371008.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-09-24
- Publication Date
- 2025-12-23
AI Technical Summary
Traditional methods for detecting hawthorn diseases are time-consuming, labor-intensive, and their accuracy depends on experience, making it difficult to achieve rapid, objective, and automated identification.
A deep convolutional neural network-based disease identification method was adopted. Through multi-view, multi-background image acquisition, manual annotation, data cleaning and enhancement, a deep convolutional neural network model was constructed for hawthorn disease identification and early warning.
It has achieved high-precision identification and automated detection of hawthorn diseases, reduced yield loss, increased the marketable fruit rate, reduced the amount of chemical pesticides used, and ensured the safe production of fruit.
Smart Images
Figure CN121190941A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of hawthorn disease control technology, and more specifically to a hawthorn disease identification method based on deep convolutional neural networks and disease images. Background Technology
[0002] In agricultural production, plant diseases are a key factor leading to decreased crop yield and deterioration in quality. Hawthorn is mainly cultivated in hilly and mountainous areas, making disease control a core aspect of its cultivation management. To promptly detect and control plant diseases, researchers have been exploring effective disease detection methods. Traditional plant disease diagnosis methods rely primarily on field surveys and experience-based judgment, which is not only time-consuming and labor-intensive but also limited in accuracy by the experience and expertise of the investigators. Intelligent plant disease identification based on convolutional neural networks has become an important research direction in precision agriculture. Therefore, how to provide a hawthorn disease identification method based on deep convolutional neural networks and disease images is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0003] In view of this, the present invention provides a hawthorn disease identification method based on deep convolutional neural networks and disease images. By systematically analyzing the features of disease images, it can achieve rapid, objective and accurate automated detection, providing technical support for early warning and precise control of hawthorn diseases.
[0004] To achieve the above objectives, the present invention provides the following technical solution:
[0005] A method for hawthorn disease identification based on deep convolutional neural networks and disease images includes the following steps:
[0006] S1. At different growth stages of hawthorn, hawthorn disease image data were collected for different parts as training samples. The image data covered multiple hawthorn disease types. A multi-view and multi-background shooting strategy was adopted when shooting.
[0007] S2. After image acquisition, healthy plants and diseased plants are distinguished by manual annotation, down to specific disease categories and severity levels.
[0008] S3. Perform data cleaning, data augmentation, and dataset expansion on manually labeled images to improve the model's generalization ability; normalize the images to unify image resolution and adapt to the model's input requirements.
[0009] S4. Construct a deep convolutional neural network model for disease identification, and train the hawthorn disease identification model based on manually labeled and preprocessed image data.
[0010] S5. Employ multi-dimensional evaluation metrics and test the detection performance of the hawthorn disease identification model using the test set in the training samples;
[0011] S6. Deploy the trained hawthorn disease recognition model to the recognition device, acquire images of the hawthorn plants to be detected, and input them into the hawthorn disease recognition model for real-time recognition.
[0012] Optionally, S2 specifically involves: using an image annotation tool to perform multi-level annotations on the image, including the health and disease status, disease type, and disease severity.
[0013] Optionally, S3 specifically involves: using the Laplacian variance algorithm to calculate the image sharpness, treating images with sharpness below a threshold as blurry images and removing them; calculating the average brightness value and pixel brightness value distribution of the image, removing images with average brightness values outside a set range and images with a pixel ratio exceeding a set threshold; enhancing and expanding the image through geometric transformation and pixel transformation, unifying the image resolution and normalizing the pixel values, and dividing the processed image into a training set, a validation set, and a test set.
[0014] Optionally, the hawthorn disease identification model includes an input module, a first convolution module, a first sampling module, a second convolution module, a second sampling module, a third convolution module, and an output module connected in sequence. The input module includes an input layer and layer convolution units, and the output module includes an average pooling layer, a fully connected layer, and a classification output layer.
[0015] Optionally, the structure of the layer convolutional unit is as follows: the input is connected to three 3×3 convolutional layers in sequence, followed by one 3×3 convolutional layer and one max pooling layer. The 3×3 convolutional layer and the 3×3 max pooling layer are connected to one splicing layer. The splicing layer is connected to two 1×1 convolutional layers, one of which is connected to one 3×3 convolutional layer. The 3×3 convolutional layer is connected to the next splicing layer. The other 1×1 convolutional layer is connected to one 7×1 convolutional layer, one 1×7 convolutional layer, one 3×3 convolutional layer, and the next splicing layer in sequence. The splicing layer is connected to the final splicing layer through one 3×3 convolutional layer and one max pooling layer.
[0016] Optionally, S5 specifically involves: inputting images from the test set into the trained hawthorn disease identification model to obtain disease identification results; comparing the identification results with the corresponding manually labeled results; calculating the AUC, accuracy, sensitivity, specificity, and F1 score of the hawthorn disease identification model; verifying the stability, evaluating the prediction efficiency, and evaluating the prediction cost of the hawthorn disease identification model; and evaluating the performance of the hawthorn disease identification model from four dimensions: core performance, stability, prediction efficiency, and prediction cost.
[0017] Optionally, S6 also includes: establishing a dynamic feedback mechanism, regularly collecting disease images, and updating model parameters through incremental learning; for different types of hawthorn plants, training corresponding hawthorn disease recognition models through transfer learning and deploying them to recognition devices.
[0018] As can be seen from the above technical solution, compared with the prior art, the present invention provides a hawthorn disease identification method based on deep convolutional neural networks and disease images, which has the following beneficial effects: The present invention utilizes the powerful feature extraction capability of the pre-trained deep convolutional neural model, combined with multi-scale data augmentation strategies and adaptive transfer learning technology, to construct an intelligent hawthorn disease identification system for complex plant image scenarios. It can achieve high-precision identification and automated and efficient detection of various diseases such as hawthorn powdery mildew, rust, and rot, while significantly improving the model's generalization ability in real complex scenarios, which can meet the real-time monitoring needs of orchards; and through early accurate identification and early warning of diseases, intelligent control of hawthorn diseases can be achieved, reducing yield loss caused by hawthorn diseases by 20%, increasing the marketable fruit rate by 15%, and reducing the amount of chemical pesticides by 20%, ensuring green and safe fruit production. Attached Figure Description
[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0020] Figure 1 This is a flowchart of the hawthorn disease identification method of the present invention;
[0021] Figure 2 This is a schematic diagram of the hawthorn disease identification model of the present invention. Detailed Implementation
[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0023] This invention discloses a method for hawthorn disease identification based on deep convolutional neural networks and disease images, such as... Figure 1 As shown, it includes the following steps:
[0024] S1. At different growth stages of hawthorn, hawthorn disease image data were collected for different parts as training samples. The image data covered multiple hawthorn disease types. A multi-view and multi-background shooting strategy was adopted when shooting.
[0025] S2. After image acquisition, healthy plants and diseased plants are distinguished by manual annotation, down to specific disease categories and severity levels.
[0026] S3. Perform data cleaning, data augmentation, and dataset expansion on manually labeled images to improve the model's generalization ability; normalize the images to unify image resolution and adapt to the model's input requirements.
[0027] S4. Construct a deep convolutional neural network model for disease identification, and train the hawthorn disease identification model based on manually labeled and preprocessed image data.
[0028] S5. Employ multi-dimensional evaluation metrics and test the detection performance of the hawthorn disease identification model using the test set in the training samples;
[0029] S6. Deploy the trained hawthorn disease recognition model to the recognition device, acquire images of the hawthorn plants to be detected, and input them into the hawthorn disease recognition model for real-time recognition.
[0030] In this embodiment of the invention, the image acquisition period covers the complete growth cycle of hawthorn (budding period, flowering period, fruit setting period, fruit enlargement period, and ripening period), the acquisition parts include key parts such as leaves (front and back), fruits, branches, and flowers, the disease types include major hawthorn diseases such as powdery mildew, rust, and rot, and the acquisition equipment uses a professional SLR camera.
[0031] The specific shooting strategy is as follows: 1. Multiple perspectives: top, bottom, and side views are captured for each sample; 2. Multiple backgrounds: natural background, solid color background, and background with varying complexity; 3. Lighting conditions: images under sunny, cloudy, shaded, and supplemental lighting conditions; 4. Shooting distance: close-up (10-30cm), medium shot (30-50cm), and distant shot (50-100cm); each plant sample should contain at least 10 images from different angles.
[0032] Furthermore, S2 specifically involves using image annotation tools to perform multi-level annotations on the image, including the health and disease status, disease type, and disease severity.
[0033] In this embodiment of the invention, the labeling tool used is Labelimg, and the severity of the disease is divided into three levels: mild, moderate, and severe.
[0034] Furthermore, S3 specifically involves: using the Laplacian variance algorithm to calculate the image sharpness, treating images with sharpness below a threshold as blurry images and removing them; calculating the average brightness value and pixel brightness value distribution of the image, removing images with average brightness values outside a set range and images with a pixel ratio exceeding a set threshold for brightness extreme values; enhancing and expanding the image through geometric transformation and pixel transformation, unifying the image resolution and normalizing the pixel values; and dividing the processed image into a training set, a validation set, and a test set.
[0035] In this embodiment of the invention, geometric transformations include random rotation, random horizontal / vertical flipping, random scaling and cropping, and affine / perspective transformations; pixel transformations include brightness adjustment, contrast adjustment, saturation adjustment, noise addition, and sharpening / blurring.
[0036] Furthermore, such as Figure 2 As shown, the hawthorn disease identification model includes an input module, a first convolution module, a first sampling module, a second convolution module, a second sampling module, a third convolution module, and an output module connected in sequence. The input module includes an input layer and a layer convolution unit, and the output module includes an average pooling layer, a fully connected layer, and a classification output layer.
[0037] Furthermore, the structure of the layer convolutional unit is as follows: the input is sequentially connected to three 3×3 convolutional layers, followed by one 3×3 convolutional layer and one max pooling layer. The 3×3 convolutional layer and the 3×3 max pooling layer are connected to one splicing layer. The splicing layer is connected to two 1×1 convolutional layers, one of which is connected to one 3×3 convolutional layer. The 3×3 convolutional layer is connected to the next splicing layer. The other 1×1 convolutional layer is sequentially connected to one 7×1 convolutional layer, one 1×7 convolutional layer, one 3×3 convolutional layer, and the next splicing layer. The splicing layer is connected to the final splicing layer through one 3×3 convolutional layer and one max pooling layer.
[0038] In this embodiment of the invention, the first convolutional module consists of four sequentially connected first convolutional units, the second convolutional module consists of seven sequentially connected second convolutional units, and the third convolutional module consists of three sequentially connected third convolutional units. The input splicing layer of the first convolutional unit is connected to one average pooling layer and three 1×1 convolutional layers, respectively. The average pooling layer is connected to the output splicing layer through one 1×1 convolutional layer, and the three 1×1 convolutional layers are connected to the output splicing layer and two 3×3 convolutional layers, respectively. One 3×3 convolutional layer is directly connected to the output splicing layer, and the other 3×3 convolutional layer is connected to the output splicing layer through one 3×3 convolutional layer.
[0039] The input splicing layer of the second convolutional unit is connected to one average pooling layer, two 1×1 convolutional layers, and one 1×7 convolutional layer. It is connected to the output splicing layer through one 1×1 convolutional layer. The average pooling layer is connected to the output splicing layer through one 1×1 convolutional layer. One of the 1×7 convolutional layers is directly connected to the output splicing layer, and the other 1×7 convolutional layer is connected to the output splicing layer through one 1×7 convolutional layer, one 7×1 convolutional layer, one 1×7 convolutional layer, and one 7×1 convolutional layer in sequence.
[0040] The input concatenation layer of the third convolutional unit is connected to one average pooling layer and three 1×1 convolutional layers. The average pooling layer is connected to the output concatenation layer through one 1×1 convolutional layer. Of the three 1×1 convolutional layers, one is directly connected to the output concatenation layer, one is connected to the output concatenation layer through one 1×3 convolutional layer and one 3×1 convolutional layer, and the last one is connected to one 1×3 convolutional layer and one 3×1 convolutional layer in sequence, and then connected to the output concatenation layer through one 1×3 convolutional layer and one 3×1 convolutional layer, respectively.
[0041] The input splicing layer of the first sampling module is connected to a 3×3 max pooling layer, a 3×3 convolutional layer and a 1×1 convolutional layer, respectively. The 3×3 max pooling layer and the 3×3 convolutional layer are directly connected to the output splicing layer. The 1×1 convolutional layer is connected to two 3×3 convolutional layers in sequence and then connected to the output splicing layer.
[0042] The input stitching layer of the second sampling module is connected to a 3×3 max pooling layer and two 1×1 convolutional layers. The 3×3 max pooling layer is directly connected to the output stitching layer. One 1×1 convolutional layer is connected to the output stitching layer through a 3×3 convolutional layer. The other 1×1 convolutional layer is connected to the output stitching layer after passing through a 1×7 convolutional layer, a 7×1 convolutional layer, and a 3×3 convolutional layer.
[0043] In this embodiment of the invention, the loss function of the hawthorn disease identification model is:
[0044] L = α·l1 + β·l2;
[0045]
[0046] In the formula, l1 represents the loss due to disease type, l2 represents the loss due to disease severity, α and β are the weights of l1 and l2 respectively, n is the sample size, and y 1,i Let y be the predicted disease type for the i-th sample. 2,i Let represent the predicted disease severity for the i-th sample. Let i be the actual disease type of the i-th sample. The true severity of the disease for the i-th sample is represented by α, and in this embodiment, α and β are 0.7 and 0.3, respectively.
[0047] Furthermore, S5 specifically involves: inputting images from the test set into the trained hawthorn disease identification model to obtain disease identification results; comparing the identification results with the corresponding manually labeled results; calculating the AUC, accuracy, sensitivity, specificity, and F1 score of the hawthorn disease identification model; verifying the stability, evaluating the prediction efficiency, and evaluating the prediction cost of the hawthorn disease identification model; and evaluating the performance of the hawthorn disease identification model from four dimensions: core performance, stability, prediction efficiency, and prediction cost.
[0048] In this embodiment of the invention, the AUC metric, accuracy, sensitivity, specificity, and F1 score are used to calculate the core performance dimension score, with weights of 0.35, 0.2, 0.2, 0.15, and 0.1, respectively. The stability verification specifically evaluates the performance differences of the model in a specific dataset (e.g., a severe disease dataset or a specific type of disease dataset). The prediction efficiency evaluation is based on the model inference time, and the prediction cost evaluation is based on the difference between the model prediction cost and the cost of manual identification. The weights of the four dimensions—core performance, stability, prediction efficiency, and prediction cost—are 0.55, 0.3, 0.1, and 0.05, respectively.
[0049] Furthermore, S6 also includes: establishing a dynamic feedback mechanism, regularly collecting disease images, and updating model parameters through incremental learning; for different types of hawthorn plants, training corresponding hawthorn disease recognition models through transfer learning and deploying them to recognition devices.
[0050] In this embodiment of the invention, transfer learning specifically refers to:
[0051] Disease image data of hawthorn plants of the target variety were collected and manually labeled and preprocessed (using the same collection and processing scheme as the original model) to obtain the target dataset. A new hawthorn disease identification model was constructed and initialized. The parameter values of the original hawthorn disease identification model were loaded into the new hawthorn disease identification model. The model hyperparameters of the backbone network were frozen. The output module in the hawthorn disease identification model was replaced with a new output module. The model was trained based on the target dataset to enable the new output module to classify and identify the target variety. Afterward, the model hyperparameters of the backbone network were unfrozen and fine-tuned. The new hawthorn disease identification model was evaluated and deployed to the corresponding identification device to identify diseases of hawthorn plants of the target variety.
[0052] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.
[0053] Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for hawthorn disease identification based on deep convolutional neural networks and disease images, characterized in that, Includes the following steps: S1. At different growth stages of hawthorn, hawthorn disease image data were collected for different parts as training samples. The image data covered multiple hawthorn disease types. A multi-view and multi-background shooting strategy was adopted when shooting. S2. After image acquisition, healthy plants and diseased plants are distinguished by manual annotation, down to specific disease categories and severity levels. S3. Perform data cleaning, data augmentation, and dataset expansion on manually labeled images to improve the model's generalization ability; normalize the images to unify image resolution and adapt to the model's input requirements. S4. Construct a deep convolutional neural network model for disease identification, and train the hawthorn disease identification model based on manually labeled and preprocessed image data. S5. Employ multi-dimensional evaluation metrics and test the detection performance of the hawthorn disease identification model using the test set in the training samples; S6. Deploy the trained hawthorn disease recognition model to the recognition device, acquire images of the hawthorn plants to be detected, and input them into the hawthorn disease recognition model for real-time recognition.
2. The hawthorn disease identification method based on deep convolutional neural networks and disease images according to claim 1, characterized in that, S2 specifically refers to: using image annotation tools to perform multi-level annotations on images, including the health and disease status, disease type, and disease severity.
3. The hawthorn disease identification method based on deep convolutional neural networks and disease images according to claim 1, characterized in that, S3 specifically involves: using the Laplacian variance algorithm to calculate the image sharpness, treating images with sharpness below a threshold as blurry images and removing them; calculating the average brightness value and pixel brightness value distribution of the image, removing images with average brightness values outside a set range and images with a pixel ratio exceeding a set threshold for brightness extreme values; enhancing and expanding the image through geometric transformation and pixel transformation, unifying the image resolution and normalizing pixel values, and dividing the processed image into training set, validation set, and test set.
4. The hawthorn disease identification method based on deep convolutional neural networks and disease images according to claim 1, characterized in that, The hawthorn disease identification model includes an input module, a first convolution module, a first sampling module, a second convolution module, a second sampling module, a third convolution module, and an output module connected in sequence. The input module includes an input layer and layer convolution units, and the output module includes an average pooling layer, a fully connected layer, and a classification output layer.
5. The hawthorn disease identification method based on deep convolutional neural network and disease images according to claim 4, characterized in that, The structure of the layer convolutional unit is as follows: the input is connected to three 3×3 convolutional layers in sequence, followed by one 3×3 convolutional layer and one max pooling layer. The 3×3 convolutional layer and the 3×3 max pooling layer are connected to one splicing layer. The splicing layer is connected to two 1×1 convolutional layers. One of the 1×1 convolutional layers is connected to one 3×3 convolutional layer. The 3×3 convolutional layer is connected to the next splicing layer. The other 1×1 convolutional layer is connected to one 7×1 convolutional layer, one 1×7 convolutional layer, one 3×3 convolutional layer, and the next splicing layer in sequence. The splicing layer is connected to the final splicing layer through one 3×3 convolutional layer and one max pooling layer.
6. The hawthorn disease identification method based on deep convolutional neural networks and disease images according to claim 1, characterized in that, S5 specifically involves: inputting images from the test set into the trained hawthorn disease identification model to obtain disease identification results; comparing the identification results with the corresponding manually labeled results; calculating the AUC, accuracy, sensitivity, specificity, and F1 score of the hawthorn disease identification model; verifying the stability, evaluating the prediction efficiency, and evaluating the prediction cost of the hawthorn disease identification model; and evaluating the performance of the hawthorn disease identification model from four dimensions: core performance, stability, prediction efficiency, and prediction cost.
7. The hawthorn disease identification method based on deep convolutional neural networks and disease images according to claim 1, characterized in that, S6 also includes: establishing a dynamic feedback mechanism, regularly collecting disease images, and updating model parameters through incremental learning; for different types of hawthorn plants, training corresponding hawthorn disease recognition models through transfer learning and deploying them to recognition devices.