Crop disease identification method based on wavelet transform and residual network fusion

Abstract

The invention relates to a crop disease identification method based on wavelet transform and residual network fusion. According to the method, a CBAM attention mechanism, wavelet transform and a residual network are fused, and a disease identification model (CropNet) for non-specified crop types is provided. According to the CropNet, firstly, Haar wavelets are utilized to perform four-stage decomposition on disease images, frequency domain features are deeply extracted, and the frequency domain features and spatial features extracted by a residual network are continuously fused; then different weights are given to the fused feature layer by using CBAM, and the attention of the model to a disease area is increased; and finally, a dual transfer learning training model is utilized to improve the accuracy and generalization of the model for identifying diseases of non-specified crop types. The identification accuracy of the CropNetA is 99.76%, the identification accuracy of the CropNetA is 99.85%, and the identification accuracy of the CropNetA is 99.86% on the PlantVillage data set, the identification accuracy of the CropNetA is 99.85% on the AI Challenger 2018 data set and the identification accuracy of the CropNetA on the self-built data set. The result shows that the method can obtain clearer and more sufficient disease characteristics while reducing the noise, improves the disease recognition precision, and provides reference for intelligent agriculture and precise recognition, prevention and control of crop diseases.

Description

Technical Field

[0001] This paper integrates the CBAM (Convolutional Block Attention Module) attention mechanism, wavelet transform (2D-DWT) and residual network (ResNeXt) to propose a crop disease recognition model (CropNet) based on the fusion of two-dimensional discrete wavelet transform and residual network. The model is trained using dual transfer learning, which enables disease recognition without specifying the crop type. It obtains clearer disease features while reducing noise and reduces the extraction of redundant features, thereby improving the model's disease recognition accuracy. Background Art

[0002] During crop growth, adverse environmental conditions such as low light, low temperatures, and high humidity can easily cause diseases, leading to reduced grain yields. According to statistics from China's Ministry of Agriculture and Rural Affairs, wheat, rice, and corn are expected to experience a resurgence of pests and diseases in crops like wheat, rice, and corn by 2024, affecting approximately 2.33 billion mu (approximately 1.3 hectares) of land and projecting yield losses exceeding 300 billion jin (approximately 1.5 million jin). Major pests and diseases in rapeseed and soybeans are expected to affect 110 million mu (approximately 1.3 hectares) of land, a 15% year-on-year increase. Therefore, accurate and effective identification of crop diseases and diagnosis of the causes of infection can enable timely and targeted prevention and control measures at the earliest stages of disease onset, preventing yield losses caused by disease spread. Traditional crop disease identification methods rely heavily on the experience of agricultural experts and manual assessment of disease characteristics. This is highly subjective, with a high rate of misjudgment and a high risk of missing the optimal opportunity for disease prevention and control. Existing deep learning and machine learning methods are mostly designed for single-crop disease identification, whereas real-world production environments often involve multiple crops. Therefore, in practical applications, when identifying unspecified crop diseases, existing methods suffer from insufficient disease information mining, limited identification of specific types, and low generalization, resulting in suboptimal disease identification accuracy. Based on 2D DWT and residual network, the present invention proposes a disease recognition model for unspecified crop types in complex backgrounds - CropNet. CropNet first continuously fuses the frequency domain features and spatial domain features extracted by two-dimensional Haar wavelet and residual network to mine more comprehensive disease features while reducing redundant features such as noise and complex background. Then, the CBAM attention mechanism is used to assign different weights to the fused features to further enhance the model's attention to disease features. Finally, the model is trained using dual transfer learning to improve the model's generalization for unspecified crop disease recognition tasks in real environments. This invention has positive significance for promoting the development of research on large-scale crop disease and unspecified crop disease recognition in complex backgrounds. Summary of the Invention

[0003] The purpose of the present invention is to provide a crop disease recognition method based on the fusion of wavelet transform and residual network, which can solve the problems of the current crop disease recognition model, such as single recognition type, insufficient disease information mining, low generalization, and incomplete extraction of key disease information, and realize effective and high-precision recognition of diseases without specifying crop types in actual environments.

[0004] To achieve the above objectives, this paper proposes a crop disease recognition method based on the fusion of a two-dimensional discrete wavelet transform and a residual network, constructing a crop-agnostic disease recognition model (CropNet). This method comprises three steps: constructing a disease dataset, building a CropNet recognition model, and finally identifying the disease. The CropNet model is based on a ResNeXt residual network. First, a two-dimensional discrete wavelet transform is introduced into the residual network to extract frequency-domain features of the disease image and continuously fuse them with deep features. A CBAM is then introduced to assign different weights to the fused features to reduce the impact of feature redundancy on crop disease recognition. Furthermore, a regularization term is introduced in the Dense layer of the CBAM to address the overfitting problem that can easily occur in CBAM. Finally, a dual transfer learning approach is used to train the model to improve its generalization. CropNet has been validated to achieve recognition accuracy rates of 99.76%, 99.85%, and 99.86% on the PlantVillage dataset, the AI ​​Challenger 2018 dataset, and self-collected kiwi, tea, apple, and grape disease datasets, respectively, exceeding existing disease classification models. It further proves that the CropNet model can obtain clearer disease features while reducing noise and reduce the extraction of redundant features, thereby improving the model's accuracy in disease recognition for unspecified crop species and its generalization in crop disease recognition tasks in actual environments.

[0005] The technical solution employed by this invention to solve the technical problem is as follows: the experimental software environment is set to TensorFlow 2.4 and Python 3.8, and the server configuration is as follows: Intel Xeon E5-2690 CPU @ 2.90GHz, 128GB of memory, and NVIDIA GeForce RTX 3080Ti. The deep learning frameworks are CUDA 11.4.0 and cuDNN 8.4.0. A CropNet model for crop-species-agnostic disease recognition is constructed by integrating neural network algorithms such as CBAM, 2D-DWT, and residual networks. This involves constructing a disease dataset, building the CropNet recognition model, and then performing crop-species-agnostic disease recognition in a real-world setting.

[0006] The disease datasets described are the PlantVillage dataset, the AI ​​Challenger 2018 dataset, and a self-built dataset. The PlantVillage dataset (https: / / github.com / spMohanty / PlantVillage-Dataset) and the AI ​​Challenger 2018 dataset (www.challenger.ai) are both public datasets. The self-built dataset is a self-collected dataset containing disease samples of kiwifruit, tea, apple, and grapes. These disease images were collected at the kiwifruit, tea, apple, and grape experimental demonstration stations of Northwest Agriculture and Forestry University, located in Mei County, Baoji City, Xixiang City, Hanzhong City, Baishui County, Weinan City, and Jingyang City, Shaanxi Province, respectively.

[0007] The recognition model CropNet is a new type of unspecified crop disease recognition network proposed by the present invention that integrates CBAM, 2D-DWT and residual network.

[0008] The beneficial effect of the present invention is that, without significantly increasing the network parameters, it effectively improves the accuracy and generalization of disease identification for unspecified crop types in real complex environments, and actively promotes the development of crop disease identification and prevention in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 Crop disease recognition process based on wavelet transform and residual network fusion Figure 2 Fusion of 2D DWT and residual network Figure 3 Four-level image decomposition via two-dimensional Haar discrete wavelet transform Figure 4 Improved fusion method of CBAM and ResNeXt101 residual module Figure 5 CBAM-DWT-ResNeXt101 network model structure Figure 6 Training process of dual transfer learning Specific implementation plan

[0010] The invention will be described in further detail below with reference to the accompanying drawings.

[0011] See also Figure 1The present invention relates to a crop disease recognition method based on the fusion of two-dimensional wavelet transform and residual network, which mainly realizes the recognition of diseases without specifying the crop species in the actual environment. The method includes three parts: disease dataset construction, recognition model CropNet construction and disease recognition. The CropNet model is based on the ResNeXt residual network. First, a two-dimensional discrete wavelet transform is introduced into the residual network to extract the frequency domain features of the disease image and continuously fuse them with the deep features. Then, CBAM is introduced to assign different weights to the fused features to reduce the impact of feature redundancy on crop disease recognition. At the same time, a regularization term is introduced on its Dense layer to address the overfitting problem that is prone to occur in CBAM. Finally, the model is trained using dual transfer learning to improve its generalization.

[0012] See also Figure 2 This method combines a four-level 2D Haar discrete wavelet decomposition with a residual network. The specific steps are as follows: Step 1: Preserving Color Features. To preserve the color characteristics of the diseased image, a first-level wavelet decomposition is performed on each of the R, G, and B channels of the input image. After reconstruction, a 3×3 convolution is performed to resize the resulting 112×112×64 feature map.

[0013] Step 2: Extraction of frequency and spatial domain features. The spatial domain features of the image are extracted from the 112×112×64 feature map using a residual block with a stride of 2, generating a 56×56×256 feature map. In the frequency domain feature extraction branch, the low-frequency components of the 112×112×64 feature map are first subjected to wavelet decomposition using a 2D DWT to extract the frequency domain features of the image, resulting in a 56×56×12 feature map. This feature map is then resized to a 56×56×256 feature map using a 1×1 convolution and a residual block with a stride of 1.

[0014] Step 3: First-layer fusion of frequency domain features and spatial domain features. The frequency domain features and spatial domain features output by the two branches in step 2 are fused and a new 56×56×256 feature map is obtained through residual connection, which serves as the input for the next stage.

[0015] Step 4: Repeat steps 2 and 3 to complete the second and third layer fusion of frequency domain features and spatial domain features, respectively, to obtain a 14×14×1024 feature map.

[0016] Step 5: The 14×14×1024 feature map generated in the above steps is subjected to the residual operation in the fourth stage to output a 7×7×2048 feature map.

[0017] See also Figure 3The above is a multi-scale feature extraction method based on 2D DWT. From left to right and from top to bottom, the first, second, third, and fourth levels of wavelet decomposition are shown. Each level of decomposition yields one low-frequency component and three high-frequency components. The method primarily uses multi-resolution analysis based on 2D DWT to focus on texture, color, and other details in the disease image, while filtering out undesirable elements such as background and noise that affect disease recognition results and weakening other interfering factors. The specific steps are as follows: Step 1: Since the image is a typical two-dimensional discrete signal, it can be decomposed into low-frequency components and high-frequency components according to equations (1) and (2).

[0018] Where, - low-frequency component, which preserves the texture, structure and other detailed information of the diseased image; h[n]- low-pass filter; ψ[k]- high-frequency component, which preserves the noise, background and other interference factors of the diseased image, including the high-frequency components of the image in the vertical, horizontal and diagonal directions; g[n]- high-pass filter.

[0019] Step 2: Use Haar wavelet transform to perform a 4-level decomposition of the disease image to extract the detailed features of the disease image and filter out interference factors such as weeds and light. The basis function of the Haar wavelet is shown in Equation (3), the corresponding scaling function is shown in Equation (4), and the filter is defined as Equation (5) and Equation (6): therefore, After transformation, we can get formula (7) and formula (8) ψ(x)=φ(2x)-φ(2x-1) (8) According to the above formula, the results of the fourth-level decomposition of the diseased sample by Haar wavelet are as follows: Figure 3 As shown in Figure 1, each decomposition results in one low-frequency component and three high-frequency components. After multiple frequency domain feature extractions, the noise is gradually suppressed and the detailed features of the disease become increasingly clear.

[0020] See also Figure 4 This improved CBAM and residual block fusion method adds CBAM to the end of the residual unit and adds a batch normalization layer before CBAM to improve the model's generalization ability, reduce the risk of overfitting, and improve recognition accuracy. Furthermore, adding CBAM within the residual unit effectively preserves the residual connections between each residual block, thereby maintaining the network's integrity in extracting disease features.

[0021] See also Figure 5This paper describes a method for constructing a CropNet crop disease recognition model. This model uses ResNeXt as its foundation and uses CBAM to assign different weights to the disease feature layer to reduce the impact of feature redundancy. The specific construction steps are as follows: Step 1: First, perform a first-level wavelet decomposition on the R, G, and B channels of the input image to retain the color features of the diseased image. The decomposed feature map is reconstructed into a 112×112×64 feature map through a 3×3 convolution operation for subsequent analysis.

[0022] Step 2: The reconstructed feature map passes through two parallel branches to extract the image's frequency and spatial domain features, respectively. In the spatial domain feature extraction branch, the feature map first uses an improved CBAM to adjust the weights of the defect layer. Then, a residual module is used to extract the image's spatial domain features, generating a 56×56×256 feature map. In the other branch, the frequency domain feature extraction branch, a 2D DWT is first used to further decompose the low-frequency components generated by the first-level wavelet decomposition, generating a 56×56×12 feature map. This feature map then undergoes 1×1 convolution to reduce its dimensionality, replacing traditional downsampling. Finally, a residual block with stride = 1 is used to resize it to a 56×56×256 feature map.

[0023] Step 3: Feature Fusion. The frequency domain features and spatial domain features output by the two branches in Step 2 are fused and weighted using an improved CBAM module. A new 56×56×256 feature map is obtained through residual connections and serves as the input for the next stage. The subsequent second and third stages operate similarly to the first stage, except that the inputs to the wavelet decomposition come from the low-frequency components of the previous stage. That is, in the subsequent two-layer frequency domain and deep feature fusion, the deep features are extracted and downsampled using different residual structures. The frequency domain features are generated by reconstructing the low-frequency information obtained from the wavelet decomposition in the previous stage through 2D DWT.

[0024] Step 5: Output. After feature fusion, a 14×14×1028 feature map is obtained, which is a fusion of the frequency domain features and spatial features obtained through the four-level decomposition and has been weighted. This feature map is input into the fourth stage, which outputs a 7×7×2048 feature map.

[0025] See also Figure 6 This multi-crop disease recognition method based on dual transfer learning is described. It primarily uses dual transfer learning to train a disease recognition model, increasing the model's accuracy and generalization for identifying diseases of unspecified crop species. CropNet, trained using dual transfer learning, effectively reduces the model's dependence on disease image samples and improves its recognition accuracy for a wide range of crop diseases. The specific steps are as follows: Step 1: Initialization: Use the ImageNet dataset as the source domain, the AI ​​Challenger 2018 dataset as the auxiliary domain, and the diseased images from the self-built dataset as the target domain.

[0026] Step 2: Pre-train the ResNeXt101 model on ImageNet to obtain the source domain model and weight 1.

[0027] Step 3: Transfer weight 1 to the improved 2D DWT-ResNeXt101 (ResNeXt101 model integrated with 2D DWT) and train it using the AI ​​Challenger 2018 dataset to obtain the auxiliary domain model and weight 2.

[0028] Step 4: Transfer weight 2 to the CBAM-2D DWT-ResNeXt model, and use the self-built dataset to train the model with fusion weight 2 to obtain the crop disease recognition model based on wavelet transform and residual network fusion - CropNet101.

Claims

1. A crop disease recognition method based on the fusion of wavelet transform and residual network. This method includes a self-constructed dataset, a crop disease recognition model (CropNet) based on the fusion of 2D Discrete Wavelet Transform (2D DWT) and residual network, and a disease recognition method based on dual transfer learning. First, disease images of four crops (apple, kiwi, tea, and grape) were collected during their peak disease season to construct a crop disease dataset. The frequency domain features of the disease extracted by the four-level 2D DWT decomposition and the spatial domain features extracted by the residual network were continuously fused to extract clear and rich disease features while mitigating the impact of undesirable components such as noise on disease recognition. Finally, dual transfer learning was used to train CropNet to improve the model's generalization for disease recognition across a wide range of crop species.

2. The self-built dataset according to claim 1, characterized in that This dataset contains 17 disease samples and 1 healthy sample of four crops, namely apple, kiwifruit, grape and tea, with a total of 32,122 disease images. These images are collected from the Baishui apple, Meixian kiwifruit, Heyang grape and Xixiang tea experimental demonstration stations in Shaanxi Province. The collection time is the peak period of various diseases. Among them, the apple disease data includes 5 diseases, including leaf drop, brown spot, mosaic, gray spot and rust, with 7,500 samples; the kiwifruit disease data includes 4 diseases, including brown spot, mosaic, anthracnose and leaf canker, with 6,115 sample images; the grape disease data includes 3 diseases, including black rot, black measles and leaf spot, with 9,393 samples; the tea disease data includes 5 diseases, including tea cake disease, ring spot, anthracnose, cloud leaf blight and tea round red spot, with 9,114 sample images.

3. CropNet, a crop disease recognition model based on the fusion of wavelet transform and residual network, features: Firstly, the two-dimensional Haar discrete wavelet transform is used to perform a four-level decomposition of the diseased image to extract frequency domain features and continuously fuse them with the spatial domain features extracted by the residual network, thereby helping the model to mine more comprehensive disease features while filtering out undesirable components such as noise that are irrelevant to the disease; then, the CBAM attention mechanism network is improved, and the improved CBAM is introduced into the residual network to increase the model's attention to the diseased area and reduce its attention to redundant features, further improving the model's disease recognition accuracy.

4. The continuous fusion method of frequency domain features and spatial domain features according to claim 3, characterized in that: First, to preserve the color features of the diseased image, the R, G, and B channels of the input 224×224×3 feature map are subjected to a first-level wavelet decomposition. After reconstruction, the feature map is adjusted to a 112×112×64 feature map through a 3×3 convolution. The feature map is then subjected to spatial domain feature extraction and frequency domain feature extraction branches respectively to obtain two 56×56×256 feature maps containing characteristic information such as lesion color, texture, and spatial structure. These two feature maps are fused and then entered into the next stage through a residual connection. After the above steps, a 7×7×2048 feature map that combines frequency domain features and spatial domain features is finally obtained.

5. The improved fusion method of CBAM and residual unit according to claim 3 is characterized in that: An improved CBAM module is added at the tail of the residual unit and before the next residual connection. At the same time, a BN layer is added before the improved CBAM to improve the generalization ability of the model, reduce the overfitting risk, and improve the recognition accuracy. At the same time, the residual connection between each residual block is retained to ensure the integrity of disease information extraction.

6. An improved CBAM according to claim 3, characterized in that L1 and L2 regularization terms are introduced into the loss function of CBMA to reduce the risk of overfitting, and the Hard-Sigmoid function is introduced to avoid the gradient vanishing phenomenon. The improved CBAM loss function is shown in formula (1), which consists of classification loss and regularization terms: （1） Where, L CBAM - CBAM loss function; L cls - Classification loss, defined as shown in formula (2); -Regularization term, in this paper by L1 Regularization term and L2 Regularization term composition; 、 -Regularization term coefficient; L1-L1 Regularization term, defined as shown in formula (3); (2) (3) (4) Where, sum ( )-sum function; y true -True value; y pred - predicted value; abs()-absolute value function; w - The parameter size of the model.

7. A crop disease identification method based on dual transfer learning, characterized by: A dual transfer learning approach was introduced to train the model. First, the ResNeXt101 model was pre-trained on ImageNet to obtain the source domain model and weight 1. Weight 1 was then transferred to a ResNeXt model fused with a 2D DWT and trained using the AI ​​Challenger 2018 dataset to obtain the auxiliary domain model and weight 2. Finally, weight 2 was transferred to a 2D DWT-ResNeXt model fused with an improved CBAM. This model, fused with weight 2, was trained using a self-built dataset to obtain a crop disease recognition method based on the fusion of wavelet transform and residual network.

Images