A method for identifying lemon yellow vein disease based on hyperspectral images and a lightweight neural network

Abstract

The present application relates to a kind of methods for identifying lemon yellow vein disease based on hyperspectral image and light neural network, comprising: collecting different wavelengths of lemon leaf hyperspectral image, processing the different wavelengths of lemon leaf hyperspectral image, inputting the processed leaf hyperspectral image into 3D-2D-LcNet model, and obtaining the identification result of lemon yellow vein disease;Wherein, the 3D-2D-LcNet model is obtained by combining light neural network ShuffleNetV2 and HybridSN.The present application can efficiently and accurately identify lemon yellow vein disease.

Description

Technical Field

[0001] The present invention relates to the technical field of lemon yellow vein disease recognition, and in particular to a lemon yellow vein disease recognition method based on hyperspectral images and lightweight neural networks. Background Art

[0002] Due to the lack of effective prevention and control agents, timely and accurate diagnosis of lemon yellow vein disease and early eradication of diseased plants have become the most important means of preventing and controlling lemon yellow vein disease.

[0003] Currently, the diagnosis of lemon yellow vein disease primarily relies on traditional manual surveys and molecular detection methods based on PCR technology. Furthermore, in natural environments, lemon plants often experience a complex phenomenon of overlapping diseases (complex diseases), with symptoms of both lemon yellow vein disease and other diseases present on the same leaf, making diagnosis of lemon yellow vein disease even more challenging. Existing diagnostic methods are cumbersome, time-consuming, and inefficient, failing to efficiently and accurately identify plants with lemon yellow vein disease.

[0004] Researchers have successfully used hyperspectral imaging technology to rapidly detect common citrus diseases such as Huanglongbing, canker, and decline, as well as other crop diseases such as tomato yellow leaf curl and apple mosaic, with satisfactory results. This suggests that hyperspectral imaging has the potential to enable efficient and accurate diagnosis of lemon yellow vein disease. Currently, there is no research on the rapid detection of lemon yellow vein disease using hyperspectral imaging technology.

[0005] Hyperspectral images contain both spectral and image information. In traditional research on plant disease detection based on hyperspectral images, the spatial information of hyperspectral images is often overlooked, and only the spectral information is used to build recognition models. In traditional plant disease recognition models, the first step is to extract spectral features, and the second step is to use a classifier for recognition. Currently, most literature adopts this traditional model. Convolutional neural networks can automatically extract the most significant features from images and are widely used in image classification and recognition. Hyperspectral images are three-dimensional hypercubes with spectral and spatial continuity. Faced with complex high-dimensional data, how to efficiently process hyperspectral images and achieve more accurate disease identification is a major difficulty in current disease diagnosis. Summary of the Invention

[0006] In order to solve the problems existing in the above-mentioned prior art, the present invention provides a lemon yellow vein disease identification method based on hyperspectral images and lightweight neural networks, which efficiently processes hyperspectral images to achieve efficient and accurate identification of lemon yellow vein disease.

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

[0008] A method for identifying lemon yellow vein disease based on hyperspectral images and lightweight neural networks, comprising:

[0009] Hyperspectral images of lemon leaves at different wavelengths are collected and processed, and the processed leaf hyperspectral images are input into a 3D-2D-LcNet model to obtain lemon yellow vein disease recognition results. The 3D-2D-LcNet model is obtained by combining the lightweight neural network ShuffleNetV2 and HybridSN.

[0010] Optionally, collecting leaf hyperspectral images at different wavelengths includes:

[0011] Hyperspectral images of lemon leaves were acquired using a hyperspectral camera at wavelengths of 402, 404, 408, 414, 423, 529, 548, 572, 681, 696, 740, 866, 958, 976, and 992 nm.

[0012] Optionally, processing the lemon leaf hyperspectral images of different wavelengths includes:

[0013] The lemon leaf hyperspectral image is scaled to a preset size, and data enhancement is performed on the scaled hyperspectral image.

[0014] Optionally, it is characterized in that the preset dimensions include: length 112×width 112×wavelength 15.

[0015] Optionally, the 3D-2D-LcNet model includes:

[0016] A 3D convolutional layer is used to extract spectral and image features from the lemon leaf hyperspectral image to obtain a three-dimensional feature map;

[0017] A function embedding layer, used to embed batch normalization (BN) and rectified linear activation functions after each 3D convolutional layer operation;

[0018] A 3D average adaptive pooling layer is used to reset the three-dimensional feature map and convert it into features suitable for 2D convolution processing;

[0019] Downsampling layer, used to downsample the converted features;

[0020] The basic layer is used to extract the features after downsampling;

[0021] A 2D convolutional layer, used to extract features from the features extracted by the base layer;

[0022] The fully connected layer is used to classify the features of the 2D convolutional layer to obtain the lemon yellow vein disease recognition result.

[0023] Optionally, the 3D convolutional layer includes:

[0024] a first 3D convolution unit, configured to extract joint features of spectrum and image from the lemon leaf hyperspectral image;

[0025] The second 3D convolution unit is used to perform secondary feature extraction on the joint feature to obtain a three-dimensional feature map.

[0026] The beneficial effects of the present invention are:

[0027] In a lemon yellow vein disease identification method based on hyperspectral images and lightweight neural networks constructed by the present invention, there is no need to perform complex operations such as collecting full-wavelength leaf hyperspectral images, segmenting regions of interest, and extracting characteristic spectra. The image and spectral fusion features can be directly extracted from the collected 15-wavelength hyperspectral images, thereby completing the identification of lemon yellow vein disease, greatly improving the recognition efficiency. Based on the lightweight neural networks ShuffleNetV2 and HybridSN, the advantages of the two are integrated to construct a lightweight neural network suitable for hyperspectral images. The final model is only 0.26MB. Compared with the ShuffleNetV2 and HybridSN convolutional neural network models, this method reduces the computational cost and improves the average recognition accuracy, which helps to achieve real-time recognition of lemon yellow vein disease on low-cost terminals. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0029] Figure 1 Schematic diagram of a basic unit and a downsampling unit according to an embodiment of the present invention;

[0030] Figure 2 Schematic diagram of a lightweight convolutional neural network according to an embodiment of the present invention;

[0031] Figure 3 Schematic diagram of confusion matrix of different models according to an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0033] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] like Figure 2 As shown, a lemon yellow vein disease recognition method based on hyperspectral image and lightweight neural network includes:

[0035] Step 1: Data collection. Six different types of lemon leaves were collected: healthy, yellow vein disease, nutrient deficiency, phytotoxicity, nutrient deficiency + yellow vein disease, and phytotoxicity + yellow vein disease. Hyperspectral images of the lemon leaves were acquired using a hyperspectral camera at wavelengths of 402, 404, 408, 414, 423, 529, 548, 572, 681, 696, 740, 866, 958, 976, and 992 nm.

[0036] Step 2: Data preprocessing and dataset construction. Scale the hyperspectral images to a uniform size of 112x112x15 (length x width x wavelength). Perform data augmentation on the scaled hyperspectral images to construct a lemon yellow vein disease hyperspectral image dataset. The dataset is randomly divided into training, validation, and test sets in a 6:2:2 ratio.

[0037] Step 3. Construct and train a lightweight neural network. Based on the lightweight neural network ShuffleNetV2 and HybridSN, the advantages of both are integrated to construct a lightweight neural network for lemon yellow vein disease hyperspectral images. The network first uses two 3x3x3 3D ​​convolution layers to extract the spectral and image features of the hyperspectral image, and embeds batch normalization (BN) and rectified linear activation function (RELU) after each 3D convolution operation; secondly, 3D average adaptive pooling is used to reset the features obtained by the 3D convolution layer and convert them into features suitable for 2D convolution processing; then, the basic units in the lightweight neural network ShuffleNetV2 are used to extract the converted features, and the number of channels of the basic units in ShuffleNetV2 is reduced and the basic units are no longer repeated, such as Figure 1As shown in the figure, this reduces computational complexity. Finally, a 1x1 convolutional layer and a Softmax classification function are used. The training data obtained in step 2 are fed into this lightweight neural network for training, resulting in a lightweight neural network model for hyperspectral imagery of jaundice. The specific training parameters are: an initial learning rate of 0.0001, a learning rate decay strategy called StepLR is used to adjust the learning rate for better convergence and performance, decaying to 50% of the previous learning rate every 20 epochs; the total number of epochs is set to 200; the batch size is set to 64; and the Adam optimizer is used.

[0038] Step 4: Disease identification: The test set images obtained in step 2 are input into the optimized convolutional neural network model obtained in step 3 to complete the identification of lemon yellow vein disease.

[0039] Specifically, lemon yellow vein disease is one of the most devastating diseases in the lemon industry. Achieving efficient and accurate diagnosis of lemon yellow vein disease is crucial for disease prevention and control and guiding field management. There have been no reports on the identification of lemon yellow vein disease based on hyperspectral imaging technology. How to effectively process hyperspectral images and obtain higher accuracy is a difficult problem faced in the rapid detection of lemon yellow vein disease. The present invention designs a lightweight neural network, which first uses two 3D convolutions to extract rich spatial and spectral joint features, and then uses the basic units in ShuffleNetV2 to extract 2D features to reduce computational complexity. The recognition performance of the 3D-2D-LcNet model proposed in the present invention and the classic convolutional neural networks 3D-ShuffleNetv2, 2D-ShuffleNetv2 and HybridSN are shown in Table 1, and the confusion matrix of the prediction set is shown in Table 1. Figure 3 As shown. The results show that compared with 3D-ShuffleNetv2 and 2D-ShuffleNetv2 using 2D or 3D convolution alone, 3D-2D-LcNet exhibits better prediction accuracy, with a prediction accuracy of 96.88%, and the computational complexity is much lower than that of 3D-ShuffleNetv2. Compared with HybridSN which uses a mixture of 3D and 2D convolutions, the 3D-2D-LcNet proposed in the present invention also performs better, with its accuracy, precision, recall rate and F1 score being much higher than HybridSN, and the model parameters and computational complexity are also much lower than HybridSN. This shows that the present invention proposes to use 3D-2D-LcNet to efficiently process hyperspectral imaging data and achieve more accurate identification of lemon yellow vein disease, which provides great help for the prevention and control of lemon yellow vein disease and field production management.

[0040] The above examples are only descriptions of the preferred modes of the present application, and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope of the present application as defined by the claims.

[0041] Table 1 Performance evaluation of different models in recognition of lemon yellow vein disease

[0042]

Claims

1. A method for identifying lemon yellow vein disease based on hyperspectral images and lightweight neural networks, characterized in that: include: Collecting hyperspectral images of lemon leaves at different wavelengths, processing the hyperspectral images of lemon leaves at different wavelengths, and inputting the processed hyperspectral images of the leaves into a 3D-2D-LcNet model to obtain lemon yellow vein disease recognition results; wherein the 3D-2D-LcNet model is obtained by combining the lightweight neural network ShuffleNetV2 and HybridSN; The 3D-2D-LcNet model includes: A 3D convolutional layer is used to extract spectral and image features from the lemon leaf hyperspectral image to obtain a three-dimensional feature map; A function embedding layer, used to embed batch normalization (BN) and rectified linear activation functions after each 3D convolutional layer operation; A 3D average adaptive pooling layer is used to reset the three-dimensional feature map and convert it into features suitable for 2D convolution processing; Downsampling layer, used to downsample the converted features; The basic layer is used to extract the features after downsampling; A 2D convolutional layer, used to extract features from the features extracted by the base layer; A fully connected layer, used to classify the features of the 2D convolutional layer to obtain the lemon yellow vein disease recognition result; The 3D convolutional layer includes: a first 3D convolution unit, configured to extract joint features of spectrum and image from the lemon leaf hyperspectral image; The second 3D convolution unit is used to perform secondary feature extraction on the joint feature to obtain a three-dimensional feature map.

2. The method for identifying lemon yellow vein disease based on hyperspectral imagery and lightweight neural network according to claim 1, characterized in that: The collection of leaf hyperspectral images at different wavelengths includes: Hyperspectral images of lemon leaves at wavelengths of 402, 404, 408, 414, 423, 529, 548, 572, 681, 696, 740, 866, 958, 976, and 992 nm were acquired using a hyperspectral camera.

3. The lemon yellow vein disease identification method based on hyperspectral image and lightweight neural network according to claim 1, characterized in that: Processing the lemon leaf hyperspectral images of different wavelengths includes: The lemon leaf hyperspectral image is scaled to a preset size, and data enhancement is performed on the scaled hyperspectral image.

4. The method for identifying lemon yellow vein disease based on hyperspectral imagery and lightweight neural network according to claim 3, characterized in that: The preset dimensions include: length 112×width 112×wavelength 15.

Images