A Method for Identifying Corn Leaf Diseases by Fusing High-Frequency Information of Images

Abstract

The present invention relates to a method for identifying maize leaf diseases by fusing high-frequency information of images, which solves the defect that it is difficult for a deep model to learn high-frequency detail information features in identifying maize leaf disease images with complex backgrounds compared with the prior art. The present invention includes the following steps: acquiring maize leaf images and performing preprocessing; constructing a maize leaf disease identification model; training the maize leaf disease identification model; acquiring maize leaf images to be identified; and obtaining the identification results of the maize leaf images. The present invention fuses the high-frequency information of the images into the lightweight MobileNetV3-Large network, and proposes a maize leaf disease identification model that fuses the high-frequency information of the images. This model has better fitting ability and improves the identification performance of maize leaf diseases in complex environments.

Description

Technical Field

[0001] The invention relates to the technical field of disease recognition, and in particular to a method for identifying corn leaf diseases by fusing high-frequency information of images. Background Art

[0002] Most corn diseases are reflected on the leaves. Traditionally, experienced agricultural experts can make corresponding disease judgments by observing the corn leaves, but this work is time-consuming and easily subject to subjective influences.

[0003] With the rapid development of computer vision and image processing technology in the agricultural field, the diagnosis level of corn leaf diseases based on digital images has been rapidly improved. Zhang et al. proposed an improved support vector machine classifier (GA-SVM), using the mean, standard deviation, area, perimeter, roundness, height, width, etc. of RGB corn leaf disease images as features to classify six types of corn leaf disease images. Xiao et al. proposed a gradient boosted decision tree (GBDT) corn leaf disease recognition method, and the recognition accuracy rate reached 92.5% on the northern leaf blight dataset. Liu et al. used the Otsu method, OpenCV morphological operations and morphological transformation methods to outline the corn leaf contour and generate masks, and combined principal component analysis and support vector machine to achieve accurate recognition of corn gray spot, corn rust, corn large spot and healthy corn leaves, with recognition accuracy rates of 90.05%, 92.64%, 91.23% and 95.78%, respectively. Traditional machine learning methods usually require manual feature extraction, which is not only time-consuming and labor-intensive, but also easy to introduce subjective bias. At the same time, since the characteristics of lesions such as shape, color, and texture of different corn leaf diseases may not differ much, it makes it very challenging to accurately extract and select features.

[0004] In recent years, DNN has developed rapidly. In this process, many researchers have designed high-accuracy corn leaf disease recognition models by improving DNN methods. Chen et al. proposed a lightweight corn leaf disease recognition model DFCANet, which mainly consists of a dual feature fusion (DFCA) module with coordinate attention and a downsampling (DS) module. The average accuracy of corn leaf disease recognition in real field environments reached 98.47%, and the model parameters and computational complexity were low. Bi et al. proposed CD-MobileNetv3 to improve the MobileNetv3 model for corn leaf disease data. The model mainly introduced the ECA module and dilated convolution, and used the bias loss function to optimize a better model. The average accuracy reached 98.23% on the mixed open source corn leaf disease dataset (CLDD). Albahli & Masood proposed an end-to-end CNNs architecture, called EANet, for accurate identification of corn leaf diseases in complex environments. The architecture improves the model's ability to focus on the diseased area by inserting the spatial channel attention mechanism into EfficientNetv2, which reduces the interference of complex background on the target area to a certain extent. Cai et al. proposed FCA-EfficientNet for corn leaf disease data in complex environments. The model mainly adds coordinate attention modules and adaptive fusion modules to EfficientNet to enhance the network's ability to focus on corn leaf disease areas and reduce background interference in recognition. The average classification accuracy of five types of corn leaf diseases and healthy leaves reached 98.78%. The above DNN methods mainly consider improvements in the network model structure, and rarely consider image spectral features to improve model performance.

[0005] In disease identification in real scenes, due to the influence of complex background on feature extraction and the complex distribution of features themselves, the anti-interference and feature extraction capabilities of the model are reduced, which leads to insufficient mining of disease information. Lin et al. converted RGB images into frequency information through DCT and introduced it into few-shot learning. They proposed a learning-based frequency selection method to select information frequency features useful for classification, and finally classified the extracted features. The classification accuracy reached about 95% on the PlantVillage dataset. Li et al. used two-dimensional discrete wavelet transform to extract four frequency features of tea disease spots and insect spots segmented by Mask R-CNN, and finally input them into a four-channel residual network for recognition, with an accuracy of 88%. Choudhary et al. used wavelets to extract frequency domain features of single-channel images and used them as model input, all of which achieved good recognition results.

[0006] However, DNN recognition models are usually easy to learn low-frequency component information during training, but difficult to fit high-frequency component information. High-frequency component information is difficult for the human eye to observe, and it contains texture details of corn leaf disease characteristics, which is very important for identifying corn leaf disease characteristics. In addition, the model's focus on high-frequency information can improve the robustness of the model to a certain extent.

[0007] Therefore, how to design a corn leaf disease recognition model that integrates high-frequency image information to identify corn leaf diseases has become a technical problem that needs to be solved urgently. Summary of the invention

[0008] The purpose of the present invention is to solve the defect in the prior art that the deep model is difficult to learn the high-frequency detail information features when identifying corn leaf disease images under complex backgrounds, and to provide a corn leaf disease recognition method that integrates image high-frequency information to solve the above problem.

[0009] In order to achieve the above object, the technical solution of the present invention is as follows:

[0010] A corn leaf disease recognition method integrating high-frequency information of an image comprises the following steps:

[0011] Obtain corn leaf images and perform preprocessing: obtain healthy and diseased corn leaf images, and perform cropping and scaling preprocessing;

[0012] Construct a corn leaf disease recognition model: Use lightweight MobileNetV3-Large as the basic network to construct a corn leaf disease recognition model;

[0013] Training of corn leaf disease recognition model: inputting the preprocessed corn leaf image into the corn leaf disease recognition model for training;

[0014] Acquisition of the image of the corn leaf to be identified: Acquisition of the image of the corn leaf to be identified and preprocessing;

[0015] Obtaining corn leaf image recognition results: Input the preprocessed corn leaf image to be identified into the trained corn leaf disease recognition model to obtain the corn leaf disease recognition result.

[0016] The construction of the corn leaf disease recognition model comprises the following steps:

[0017] The corn leaf disease recognition model is set to include a lightweight MobileNetV3-Large, a high-frequency filter, and a high-frequency feature extraction module, wherein the lightweight MobileNetV3-Large is used as the basic network of the corn leaf disease recognition model;

[0018] Setting up lightweight MobileNetV3-Large is divided into three parts:

[0019] The first part consists of a convolutional layer, which extracts features and compresses dimensions through 3×3 convolution;

[0020] The second part contains 15 Inverted bottleneck blocks, of which the 1st to 3rd and 7th to 12th layers use depth-wise separable convolutions with 3×3 convolution kernels, and the remaining Inverted bottleneck blocks use depth-wise separable convolutions with 5×5 convolution kernels, which improves model performance while reducing the number of model parameters.

[0021] The third part uses a 1×1 convolution layer to expand the number of feature channels, then uses an adaptive average pooling layer to convert it into a feature vector, and finally inputs it into a classifier to output the category;

[0022] The first part is shallow convolution, and the second part is deep convolution. Shallow convolution extracts rich low-frequency semantic information of the image, and deep convolution extracts abstract information.

[0023] A high-frequency feature extraction module is inserted after the 4th, 7th, and 11th Inverted bottleneck blocks of the lightweight MobileNetV3-Large to obtain high-frequency feature information and compress the image size.

[0024] The high-frequency filter is set to use the Gaussian high-pass filtering method to filter out abundant low-frequency information while retaining high-frequency detail information that is difficult to fit;

[0025] A high-frequency feature extraction module is set up, which consists of adaptive average pooling, 1×1 point convolution, 3×3 convolution and 1×1 point convolution. An adaptive ACON-C activation function is used between convolutions to extract texture information features from high-frequency detail information. At the same time, the high-frequency detail information of different spatial sizes is connected with the corn leaf disease features extracted by MobileNetV3-Large.

[0026] The training of the corn leaf disease recognition model comprises the following steps:

[0027] The preprocessed corn leaf disease image is resized to 224×224 and input into the lightweight MobileNetV3-Large to obtain the category probability;

[0028] The obtained category probability and true label are input into the cross entropy loss function to obtain the loss value and the weight of the lightweight MobileNetV3-Large is reversely optimized through the stochastic gradient descent method, and the trained lightweight MobileNetV3-Large weight parameters are fixed; the cross entropy loss value is calculated as follows:

[0029]

[0030] Among them, B represents the number of corn leaf disease images input in each batch, f represents the MobileNetV3-Large model, and y i represents the true label value of the i-th image data, x i represents the i-th input image data, L f The loss value calculated for the MobileNetV3-Large model;

[0031] The corn leaf disease image is resized to 224×224 and the high-frequency texture information of the image is extracted through three high-frequency filters. The cutoff frequencies of the three high-frequency filters are 5, 20, and 50 respectively. The Gaussian filter function formula is as follows:

[0032]

[0033] Among them, u and v are the frequency domain position coordinates, H and W are the height and width of the input feature, σ is the cutoff frequency, and H(u,v) is a Gaussian high-pass filter;

[0034] The high-frequency texture information obtained by the high-frequency filter with a cutoff frequency of 5 is input into the high-frequency feature extraction module after the fourth Inverted bottleneck block of the lightweight MobileNetV3-Large to obtain the frequency feature information. At the same time, the corn leaf disease image is resized to 224×224 and input into the trained lightweight MobileNetV3-Large to obtain the fourth layer Inverted bottleneck block and then output the semantic features. The frequency feature information and the semantic features are added and fused to obtain the fused feature A.

[0035] The fused feature A is input into the 5th to 7th layer Inverted bottleneck block to extract semantic features. At the same time, the high-frequency texture information obtained by the high-frequency filter with a cutoff frequency of 20 is input into the high-frequency feature extraction module after the 7th Inverted bottleneck block to obtain the frequency feature, which is then added and fused with the semantic feature to obtain the fused feature B.

[0036] The fused feature B is input into the 8th to 10th layer Inverted bottleneck block to extract semantic features. At the same time, the high-frequency texture information obtained by the high-frequency filter with a cutoff frequency of 50 is input into the high-frequency feature extraction module after the 11th Inverted bottleneck block to obtain the frequency feature, which is then added and fused with the semantic feature to obtain the fused feature C.

[0037] The fused feature C is input into the remaining 5 Inverted bottleneck blocks to output the final abstract semantic features and the feature vector is obtained using the adaptive global average pooling layer. Finally, the feature vector is input into the fully connected layer classifier to obtain the category probability.

[0038] The cross entropy loss function is used to fine-tune the parameter values ​​in the above three high-frequency feature extraction modules, and the best recognition overall model weights are saved for testing. The cross entropy loss value is calculated as follows:

[0039]

[0040] Where B represents the number of corn leaf disease images input in each batch, π represents the corn leaf disease recognition model, and y i represents the true label value of the i-th image data, x i represents the i-th input image data, L π Loss value calculated by the corn leaf disease recognition model.

[0041] Beneficial Effects

[0042] The present invention discloses a method for identifying corn leaf diseases by integrating high-frequency information of images. Compared with the prior art, the method integrates the high-frequency information of images into a lightweight MobileNetV3-Large network, and proposes a corn leaf disease identification model by integrating high-frequency information of images. The model has better fitting ability and improves the recognition performance of corn leaf diseases in complex environments.

[0043] The present invention uses MobileNetV3-Large as the backbone network, and improves the model's ability to fit high-frequency information by adding high-frequency component information of corn leaf disease images to the last three layers of MobileNetV3-Large, thereby improving the model's generalization performance; through a high-frequency feature extraction module (HFFE), it connects layers of different spatial sizes; and by introducing an ACON-C activation function in the high-frequency feature extraction block, the nonlinear fitting ability of high-frequency information features is improved. In addition, the corn leaf disease recognition model constructed by the present invention is an end-to-end one, which eliminates the need to independently extract high-frequency information components of the input image.

[0044] A large number of experiments show that the method proposed by the present invention has the advantages of high precision and strong robustness. On the public dataset, the method of the present invention can still improve the recognition accuracy of the basic model MobileNetV3-Large. In addition, after adding noise to the test images of the maize leaf disease dataset obtained in this study, the method of the present invention has higher recognition accuracy and strong robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is the sequence diagram of the method of the present invention;

[0046] Figure 2 It is the structural schematic diagram of the maize leaf disease recognition model described in the present invention;

[0047] Figure 3 It is the structural schematic diagram of the high-frequency feature extraction module described in the present invention;

[0048] Figure 4 It is the comparison chart of the accuracy values of the models trained by adding different high-frequency information involved in the present invention;

[0049] Figure 5 It is the comparison result chart of the recognition accuracy under different noise levels involved in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0050] In order to have a further understanding and recognition of the structural features and achieved effects of the present invention, the following is a detailed description with the help of preferred embodiments and drawings:

[0051] As Figure 1 shown, a maize leaf disease recognition method integrating high-frequency information of images according to the present invention includes the following steps:

[0052] The first step is to obtain maize leaf images and perform preprocessing: obtain healthy and diseased images of maize leaves, and perform preprocessing such as cropping and scaling.

[0053] The second step is to construct a maize leaf disease recognition model: use the lightweight MobileNetV3-Large as the basic network to construct a maize leaf disease recognition model. In order to improve the model's learning ability for high-frequency information, the present invention uses the lightweight MobileNetV3-Large as the basic network, and inserts the high-frequency feature information obtained by the high-frequency filter into the last three layers of this network in sequence.

[0054] As Figure 2As shown in the figure. For the original input image, it is first input into MobileNetV3-Large for feature extraction. Then, different high-frequency information of the image is obtained and standardized in turn through three high-frequency filter modules with different cutoff frequencies σ. Here, we choose the cutoff frequency of σ = [5, 20, 50] to obtain the high-frequency information of the corn leaf disease image as the input of the network model. Considering that the first few layers of the deep neural network extract the shallow semantic features of the image, a lower cutoff frequency can obtain more low-level features to enrich the shallow semantic information; as the network level goes deeper, the DNN begins to combine and abstract these low-level features to form a higher-level feature representation. These deep features are more abstract and can express the global information of the image, but it is difficult to learn the high-frequency details of the image. Therefore, high-frequency information with a higher cutoff frequency is added to improve feature diversity. Secondly, the high-frequency information obtained is extracted through the high-frequency feature extraction block (HFFE block), and the features are compressed to the same output dimension as the output layer of the last three layers of MobileNetV3-Large. Then, the high-frequency features and image information features with the same dimension are added and fused. Finally, the features extracted from the last layer of the network model are input into the classifier for classification.

[0055] The construction of corn leaf disease recognition model includes the following steps:

[0056] (1) The corn leaf disease recognition model is set to include a lightweight MobileNetV3-Large, a high-frequency filter, and a high-frequency feature extraction module, among which the lightweight MobileNetV3-Large is used as the basic network of the corn leaf disease recognition model.

[0057] (2) Set the lightweight MobileNetV3-Large to be divided into three parts:

[0058] The first part consists of a convolutional layer, which extracts features and compresses dimensions through 3×3 convolution;

[0059] The second part contains 15 Inverted bottleneck blocks, of which the 1st to 3rd and 7th to 12th layers use depth-wise separable convolutions with 3×3 convolution kernels, and the remaining Inverted bottleneck blocks use depth-wise separable convolutions with 5×5 convolution kernels, which improves model performance while reducing the number of model parameters.

[0060] The third part uses a 1×1 convolution layer to expand the number of feature channels, then uses an adaptive average pooling layer to convert it into a feature vector, and finally inputs it into a classifier to output the category;

[0061] The first part is shallow convolution, and the second part is deep convolution. Shallow convolution extracts rich low-frequency semantic information of the image, and deep convolution extracts abstract information.

[0062] (3) High-frequency feature extraction modules are inserted after the 4th, 7th, and 11th Inverted bottleneck blocks of the lightweight MobileNetV3-Large to obtain high-frequency feature information and compress the image size.

[0063] (4) The high-frequency filter is set to use the Gaussian high-pass filtering method to filter out the rich low-frequency information while retaining the high-frequency detail information that is difficult to fit.

[0064] In order to obtain the high-frequency component information of the corn leaf disease image, the present invention uses a Gaussian high-pass filter. First, a two-dimensional corn leaf disease image can be expressed as f(x, y), and its Fourier transform formula is as follows:

[0065]

[0066] Among them, M is the height of the input image, N is the width of the input image, f(x,y) is the time domain image, x and y are the time domain position coordinates, F(u,v) is the frequency domain image, u and v are the frequency domain position coordinates.

[0067] In order to filter out the rich low-frequency information and retain the high-frequency detail information, a Gaussian high-pass filter is used for filtering. The high-pass filter function is as follows:

[0068]

[0069] Where σ is the cutoff frequency, and H(u,v) is the Gaussian high-pass filter. The right-hand side of the equation is the Gaussian low-pass filter. The Gaussian high-pass filter of equation (2) is obtained by subtracting the Gaussian low-pass filter from the all-pass filter 1.

[0070] Then, the obtained image frequency information is multiplied by the Gaussian high-pass filter at the corresponding position to obtain the image high-frequency information, as follows:

[0071] F′(u,v)=F(u,v)·H(u,v)

[0072] Where F′(u,v) is the high frequency image after filtering.

[0073] Finally, the high-frequency image is inversely Fourier transformed to obtain the filtered high-frequency time domain image, as follows:

[0074]

[0075] Where f′(x,y) is the high-frequency image after filtering.

[0076] In the present invention, by controlling the cutoff frequency σ, images with different high-frequency information are extracted from the corn leaf disease image, which are used for subsequent input into the network model for high-frequency feature extraction, thereby improving the network model to learn richer corn leaf disease edge texture detail information.

[0077] (5) A high-frequency feature extraction module is set up, which is composed of adaptive average pooling, 1×1 point convolution, 3×3 convolution and 1×1 point convolution. An adaptive ACON-C activation function is used between convolutions to extract texture information features from high-frequency detail information. At the same time, the high-frequency detail information of different spatial sizes is connected with the corn leaf disease features extracted by MobileNetV3-Large.

[0078] Generally speaking, in the corn leaf disease classification network, the layer space size gradually decreases, so the high-frequency component information cannot be directly used. In order to solve this limitation, the present invention designs a high-frequency feature extraction block HFFE, which can connect corn leaf disease features of different spatial sizes. Figure 3 As shown in the figure, HFFE consists of pooling operations and convolutions. First, in order to eliminate a large amount of redundant information in the high-frequency component image and retain key information while reducing the image space size, an adaptive average pooling is used, and then high-frequency features are extracted through inverted bottleneck convolution blocks consisting of 1×1, 3×3, and 1×1. Batch normalization (BN) and ACON-C activation function are used between convolutional layers for nonlinear processing, and ACON-C is not used in the last layer.

[0079] The third step is to train the corn leaf disease recognition model: input the preprocessed corn leaf image into the corn leaf disease recognition model for training.

[0080] (1) The preprocessed corn leaf disease image is resized to 224×224 and input into the lightweight MobileNetV3-Large to obtain the category probability.

[0081] (2) Input the obtained category probability and true label into the cross entropy loss function to obtain the loss value and reversely optimize the weight of lightweight MobileNetV3-Large through the stochastic gradient descent method, and fix the trained lightweight MobileNetV3-Large weight parameters; the cross entropy loss value is calculated as follows:

[0082]

[0083] Among them, B represents the number of corn leaf disease images input in each batch, f represents the MobileNetV3-Large model, and y irepresents the true label value of the i-th image data, x i represents the i-th input image data, L f Loss value calculated for the MobileNetV3-Large model.

[0084] (3) The corn leaf disease image is resized to 224×224 and the high-frequency texture information of the image is extracted through three high-frequency filters. The cutoff frequencies of the three high-frequency filters are 5, 20, and 50, respectively. The Gaussian filter function formula is as follows:

[0085]

[0086] Among them, u and v are the frequency domain position coordinates, H and W are the height and width of the input feature, σ is the cutoff frequency, and H(u,v) is a Gaussian high-pass filter.

[0087] (4) The high-frequency texture information obtained by the high-frequency filter with a cutoff frequency of 5 is input into the high-frequency feature extraction module after the fourth Inverted bottleneck block of the lightweight MobileNetV3-Large to obtain the frequency feature information. At the same time, the corn leaf disease image is resized to 224×224 and input into the trained lightweight MobileNetV3-Large to obtain the fourth-layer Inverted bottleneck block and then output the semantic features. The frequency feature information and the semantic features are added and fused to obtain the fused feature A.

[0088] (5) The fused feature A is input into the 5th to 7th layer Inverted bottleneck block to extract the semantic feature. At the same time, the high-frequency texture information obtained by the high-frequency filter with a cutoff frequency of 20 is input into the high-frequency feature extraction module after the 7th Inverted bottleneck block to obtain the frequency feature, which is then added and fused with the semantic feature to obtain the fused feature B.

[0089] (6) The fused feature B is input into the 8th to 10th layer Inverted bottleneck block to extract semantic features. At the same time, the high-frequency texture information obtained by the high-frequency filter with a cutoff frequency of 50 is input into the high-frequency feature extraction module after the 11th Inverted bottleneck block to obtain the frequency feature, which is then added and fused with the semantic feature to obtain the fused feature C.

[0090] (7) The fused feature C is input into the remaining five Inverted bottleneck blocks to output the final abstract semantic features and the adaptive global average pooling layer is used to obtain the feature vector. Finally, the feature vector is input into the fully connected layer classifier to obtain the category probability.

[0091] The cross entropy loss function is used to fine-tune the parameter values ​​in the above three high-frequency feature extraction modules, and the best recognition overall model weights are saved for testing. The cross entropy loss value is calculated as follows:

[0092]

[0093] Where B represents the number of corn leaf disease images input in each batch, π represents the corn leaf disease recognition model, and y i represents the true label value of the i-th image data, x i represents the i-th input image data, L π The loss value calculated for the overall model.

[0094] Through the above two stages of training, the model's ability to fit high-frequency information can be improved. In addition, the second stage of training only adjusts the parameters of the high-frequency feature extraction block, which can greatly improve the training efficiency.

[0095] The fourth step is to obtain the image of the corn leaf to be identified: obtain the image of the corn leaf to be identified and perform preprocessing.

[0096] The fifth step is to obtain the corn leaf image recognition result: the pre-processed corn leaf image to be identified is input into the trained corn leaf disease recognition model to obtain the corn leaf disease recognition result.

[0097] In order to verify the recognition accuracy of the method proposed in the present invention, the data of four common types of corn leaf diseases and healthy leaves obtained by the present invention were experimentally compared. The experimental results are shown in Table 1. As can be seen from the table, the method proposed in the present invention is significantly better than all the compared methods. Compared with the basic model MobileNetV3-Large, it is improved by 2.0%. Compared with Transformer, the model of the present invention performs best in Accuracy, Precision, Recall and F1-score. It may be due to the large number of Transformer parameters that the model is prone to overfitting, affecting the recognition accuracy.

[0098] Table 1. Comparison of test results of corn leaf disease recognition model

[0099]

[0100] In order to demonstrate the recognition performance of the benchmark model MobileNetV3-Large by the high-frequency feature information used in the present invention, a comparative experiment was conducted on the recognition model of the high-frequency information of the last three layers of the benchmark model with fusion cutoff frequencies of 5, 20 and 50. Figure 4 As shown. Figure 4 It can be seen that the models that incorporate high-frequency information have improved in accuracy, which shows the effectiveness of adding high-frequency information of corn leaf diseases. When the cutoff frequency f = [5, 20, 50] is used to build the model, the recognition accuracy is the highest, reaching 0.957. However, when the same cutoff frequency is used to obtain high-frequency information as the input of the last three layers of the model, the accuracy decreases compared with the combination of f = [5, 20, 50]. We speculate that this may be because the deeper the model, the more serious the loss of high-frequency information, and the weaker the ability to extract high-frequency information features, resulting in a decrease in the model's fitting ability, resulting in relatively low recognition accuracy. In addition, when a lower cutoff frequency is used to obtain high-frequency information of corn leaf diseases as input, the accuracy will be higher, which may be because the model can better fit the low-frequency information, thereby improving the recognition accuracy.

[0101] In order to evaluate the robustness of the network proposed in the present invention, different degrees of Gaussian and salt-and-pepper noise were added to the data of four common types of corn leaf diseases and healthy leaves obtained by the present invention to evaluate the recognition accuracy of our network. Figure 5 As shown. The horizontal axis and the vertical axis represent the added noise intensity and the recognition accuracy, respectively. Obviously, with the increase of noise intensity, the accuracy of various recognition methods decreases, because it is difficult to extract effective features from images with noise interference, thus affecting the recognition accuracy. In addition, under different noise types and noise levels, the accuracy of the method proposed in the present invention is higher than that of the basic network MobileNetV3-Large. With the increase of noise intensity, the recognition accuracy of the MobileNetV3-Large network decreases faster than that of the method proposed in the present invention. Therefore, the method of the present invention has better performance in the accuracy and robustness of the acquired corn leaf disease data.

[0102] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions only describe the principles of the present invention. The present invention may be subject to various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements fall within the scope of the present invention. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A corn leaf disease recognition method integrating high-frequency information of an image, characterized in that: The following steps are involved: 11) Obtaining corn leaf images and preprocessing: Obtaining healthy and diseased corn leaf images, and performing cropping and scaling preprocessing; 12) Construct a corn leaf disease recognition model: Use lightweight MobileNetV3-Large as the basic network to construct a corn leaf disease recognition model; The construction of the corn leaf disease recognition model comprises the following steps: 121) The corn leaf disease recognition model is set to include a lightweight MobileNetV3-Large, a high-frequency filter and a high-frequency feature extraction module, wherein the lightweight MobileNetV3-Large is used as the basic network of the corn leaf disease recognition model; 122) Set the lightweight MobileNetV3-Large to be divided into three parts: The first part consists of a convolutional layer, which extracts features and compresses dimensions through 3×3 convolution; The second part contains 15 Inverted bottleneck blocks, of which the 1st to 3rd and 7th to 12th layers use depth-wise separable convolutions with 3×3 convolution kernels, and the remaining Inverted bottleneck blocks use depth-wise separable convolutions with 5×5 convolution kernels, which improves model performance while reducing the number of model parameters. The third part uses a 1×1 convolution layer to expand the number of feature channels, then uses an adaptive average pooling layer to convert it into a feature vector, and finally inputs it into a classifier to output the category; The first part is shallow convolution, and the second part is deep convolution. Shallow convolution extracts rich low-frequency semantic information of the image, and deep convolution extracts abstract information. 123) Insert high-frequency feature extraction modules after the 4th, 7th and 11th Inverted bottleneck blocks of the lightweight MobileNetV3-Large to obtain high-frequency feature information and compress the image size; 124) The high-frequency filter is set to adopt the Gaussian high-pass filtering method to filter out the abundant low-frequency information and retain the high-frequency detail information that is difficult to fit; 125) Setting a high-frequency feature extraction module, which is composed of adaptive average pooling, 1×1 point convolution, 3×3 convolution and 1×1 point convolution, and using an adaptive ACON-C activation function between convolutions to extract texture information features from high-frequency detail information. At the same time, high-frequency detail information of different spatial sizes is connected with the corn leaf disease features extracted by MobileNetV3-Large; 13) Training of corn leaf disease recognition model: inputting the preprocessed corn leaf image into the corn leaf disease recognition model for training; 14) Acquiring the image of the corn leaf to be identified: acquiring the image of the corn leaf to be identified and performing preprocessing; 15) Obtaining corn leaf image recognition results: The preprocessed corn leaf image to be identified is input into the trained corn leaf disease recognition model to obtain the corn leaf disease recognition result.

2. The corn leaf disease identification method of integrating high-frequency information of images according to claim 1 is characterized in that: The training of the corn leaf disease recognition model comprises the following steps: 21) Resize the preprocessed corn leaf disease image to 224×224 and input it into the lightweight MobileNetV3-Large to obtain the category probability; 22) Input the obtained category probability and true label into the cross entropy loss function to obtain the loss value and reversely optimize the weight of lightweight MobileNetV3-Large through the stochastic gradient descent method, and fix the trained lightweight MobileNetV3-Large weight parameters; the cross entropy loss value is calculated as follows: Among them, B represents the number of corn leaf disease images input in each batch, f represents the MobileNetV3-Large model, and y i represents the true label value of the i-th image data, x i represents the i-th input image data, L f The loss value calculated for the MobileNetV3-Large model; 23) The corn leaf disease image is resized to 224×224 and the high-frequency texture information of the image is extracted through three high-frequency filters. The cutoff frequencies of the three high-frequency filters are 5, 20, and 50 respectively. The Gaussian filter function formula is as follows: Among them, u and v are the frequency domain position coordinates, H and W are the height and width of the input feature, σ is the cutoff frequency, and H(u,v) is a Gaussian high-pass filter; 24) The high-frequency texture information obtained by the high-frequency filter with a cutoff frequency of 5 is input into the high-frequency feature extraction module after the 4th Inverted bottleneck block of the lightweight MobileNetV3-Large to obtain the frequency feature information. At the same time, the corn leaf disease image is resized to 224×224 and then input into the trained lightweight MobileNetV3-Large to obtain the 4th layer Inverted bottleneck block and then output the semantic features. The frequency feature information and the semantic features are added and fused to obtain the fused feature A; 25) The fused feature A is input into the 5th to 7th layer Inverted bottleneck block to extract semantic features. At the same time, the high-frequency texture information obtained by the high-frequency filter with a cutoff frequency of 20 is input into the high-frequency feature extraction module after the 7th Inverted bottleneck block to obtain the frequency feature, and then added and fused with the semantic feature to obtain the fused feature B; 26) The fused feature B is input into the 8th to 10th layer Inverted bottleneck block to extract semantic features. At the same time, the high-frequency texture information obtained by the high-frequency filter with a cutoff frequency of 50 is input into the high-frequency feature extraction module after the 11th Inverted bottleneck block to obtain the frequency feature, and then added and fused with the semantic feature to obtain the fused feature C; 27) Input the fusion feature C into the remaining 5 Inverted bottleneck blocks to output the final abstract semantic features and use the adaptive global average pooling layer to obtain the feature vector. Finally, input the feature vector into the fully connected layer classifier to obtain the category probability. The cross entropy loss function is used to fine-tune the parameter values ​​in the above three high-frequency feature extraction modules, and the best recognition overall model weights are saved for testing. The cross entropy loss value is calculated as follows: Where B represents the number of corn leaf disease images input in each batch, π represents the corn leaf disease recognition model, and y i represents the true label value of the i-th image data, x i represents the i-th input image data, L π Loss value calculated by the corn leaf disease recognition model.

Images