CN112861752A

CN112861752A - Crop disease identification method and system based on DCGAN and RDN

Info

Publication number: CN112861752A
Application number: CN202110201803.3A
Authority: CN
Inventors: 周长建; 宋佳; 邢金阁; 周思寒; 冯宝龙; 刘宇航
Original assignee: Northeast Agricultural University
Current assignee: Northeast Agricultural University
Priority date: 2021-02-23
Filing date: 2021-02-23
Publication date: 2021-05-28
Anticipated expiration: 2041-02-23
Also published as: CN112861752B

Abstract

The invention relates to a crop disease identification method and system based on DCGAN and RDN, belonging to the field of agricultural informatization and plant protection, and the method comprises the following steps: firstly, data collection is carried out, wherein the data collection comprises a network public data set and manually collected data, and then the accuracy and the distribution balance of a training data set are guaranteed by utilizing technologies such as data visualization, data cleaning, DCGAN data generation and the like; dividing the processed data according to the proportion of a 60% training set, a 20% verification set and a 20% testing set; the method comprises the steps of constructing a deep residual error network (RDN) recognition model, setting model training parameters, and loading a training set and a verification set for model training; and finally, applying the trained model to a crop leaf disease recognition system for crop disease prediction, and returning the crop disease category and probability to the system. The method can identify various diseases of various crops, and particularly can improve the identification accuracy under the condition that the samples are not uniformly distributed.

Description

Crop disease identification method and system based on DCGAN and RDN

Technical Field

The invention relates to the field of agricultural informatization and plant protection, in particular to a crop disease identification method and system based on the combination of DCGAN and RDN. In particular to an identification method and an identification system which are used for enhancing data by using a DCGAN technology under the condition of large-scale crop leaf training data distribution imbalance and improving detection accuracy by combining an RDN algorithm.

Background

1. The term of art:

(1) DCGAN. DCGAN is an abbreviation of Deep Convolutional antagonistic generation network (Deep Convolutional adaptive Networks), and the invention mainly utilizes DCGAN technology to enhance data and is used for expanding training data so as to enhance the generalization capability of the Deep learning network and increase the recognition accuracy.

(2) Deep Residual Network (Deep Residual Network). A deep learning model, please refer to fig. 1, which is proposed by (He, k., Zhang, x., Ren, s., & Sun, j.,2016), and its main idea is to directly pass input x to output as an initial result by means of residual Connections (short Connections), and the output result is h (x) ═ f (x) + x, and when f (x) ═ 0, then h (x) ═ x, i.e. the above mentioned identity mapping. The deep residual network is equivalent to changing the learning objective, and does not learn a complete output, but the difference between the target values h (x) and x, i.e. the so-called residual f (x) ═ h (x) -x, so that the following training objective is to approach the residual result to 0, and the accuracy does not decrease as the network deepens. The method solves the problem that the model can not be converged due to the fact that gradient dispersion/explosion becomes an obstacle of a network in a training deep level caused by the deepening of the training level.

(3) Deep Dense connection Network (Deep Dense Network). Referring to fig. 2, a deep learning model is proposed by (Huang, g., Liu, z., Laurens, v.d.m.et al.,2017), and the core idea is to directly connect all layers in a network on the premise of ensuring the maximum information transmission between the layers. In a conventional convolutional neural Network, if there are L layers, there are L connections, but in a sense Network, there are L (L +1)/2 connections, i.e., the input of each layer is the output from all the previous layers. This approach has several advantages: a) the problem of gradient disappearance is alleviated; b) the transmission characteristic strength of the upper layer and the lower layer is increased, so that the characteristics can be more effectively utilized; c) the number of parameters is reduced, and the calculation efficiency is improved.

(4) RDN is used. RDN is an abbreviation for Residual Dense connection Network (Residual Dense Network). The method carries out crop disease identification by constructing the RDN deep learning network. The residual dense connection network is firstly proposed by (Zhang, y., Tian, y., Kong, y., Zhong, b., & Fu, y.2018), and is mainly applied to image super-resolution work. The network combines the advantages of the deep dense connection network and the deep dense connection network, and is combined into a new deep learning network, so that the problem of gradient dispersion/explosion is avoided, and the calculation efficiency is further improved, please refer to fig. 3.

2. Analysis of current research situation at home and abroad

At present, mainstream crop disease identification methods are mainly divided into traditional computer vision and machine learning identification methods and deep learning-based identification methods. Nanehkaran, YA et al, (2020) proposed a disease detection method, in which the author first segmented the image and then detected it, and finally the average detection accuracy reached 75.59%. Yuan, L, Yan, P, Han, WY.et al (2019) provides a tea tree anthracnose detection method, an author firstly binarizes an image, sets a threshold value for image segmentation, and then classifies the image by using an unsupervised method, and the accuracy of the method on a tea tree leaf anthracnose data set can reach 94%.

Most of the traditional computer vision methods need to manually extract image features and then classify and identify the features. The method needs strong prior knowledge, has very high requirements on human experience, and has limited characteristics due to manual selection, so that the identification accuracy rate is difficult to reach the actual application standard; moreover, due to the various crop diseases, manual labeling requires a large amount of labor, and the labor cost is very high. The deep learning technology can utilize a deep convolutional neural network to automatically extract and classify features, the deep learning model has certain learning capacity by setting an optimization function and a loss function, and stronger robust description can be generated by training on a large number of data sets, so that very high recognition accuracy is achieved.

Zhong, y, Zhao, m. (2020) proposes an apple leaf disease identification model, and authors propose a classification method of one but a plurality of classes of labels by comparing superiority and inferiority of learning at different depths, with the highest accuracy reaching 93.71%. Ferentinos, K.P, (2018) proposes a crop disease identification method of a deep convolutional network, and authors classify crop leaf diseases into healthy leaves and diseased leaves, and the accuracy can reach 99.53% at most. Jiang, F, Lu, Y, Chen, Y.et al (2020) proposes a rice leaf disease recognition model based on deep learning and SVM (support vector machine) combination, and the method can be divided into two parts, firstly, CNN is used for extracting features, and then the features are put into SVM for classification, and finally the accuracy reaches 96.8%. Chen, j, Zhang, d.et al. (2020) proposes a crop disease identification method based on transfer learning, and authors perform disease identification by using a model trained on ImageNet, and finally reach an accuracy of 91.83. Picon, A., Alvarez-Gila, A., Seitz, M.et al. (2019) provides a wheat leaf disease identification method of a self-adaptive deep residual error network, and the final accuracy rate reaches 87%.

The above methods all achieve good effects in experimental environments, but have certain disadvantages, mainly including:

(1) most of published results at present are experimental data on a small data set, mainly focusing on a crop for research, and the other results only divide the data into healthy leaves and diseased leaves for classification and identification, which may obtain higher accuracy, but lack universality and have low practical application value.

(2) Data distribution is not considered, when the data volume is large and the types of diseases are more, common diseases may occur, the data volume is very large, and some unusual diseases cause unbalanced training data distribution due to reasons such as large data acquisition difficulty and small data volume, so that the use effect is influenced under the condition.

(3) At present, most methods multiplex the existing models, and the model innovativeness is low.

(4) Patent document CN110929610A (application number: CN201911101895.7) discloses a method for identifying crop diseases based on CNN model and transfer learning, which first performs enhancement processing on data by using traditional data enhancement techniques (such as rotation transformation, translation transformation, light transformation, etc.), and then performs crop disease identification by using inclusion-v 3 model and MobileNet model trained on ImageNet. The method has two main disadvantages: a) experiments show that the traditional data enhancement technology mainly performs size, angle and other transformations on the original image, the original bottom layer characteristics of the image are not transformed, the accuracy is not greatly improved, and the generalization capability of the model cannot be expressed; b) the precondition for performing migration training by using the trained model on ImageNet is that the recognized data type is similar to or consistent with that of ImageNet, according to ImageNet official network (http://www.image-net.org/) Description, the data set is various in types and large in data, but the data related to crop diseases are rarely available, so that the method cannot have significant guiding significance for crop disease identification.

(5) Patent document CN111709481A (application number: CN202010555276.1) discloses a tobacco disease identification method, system, platform and storage medium, which mainly uses a traditional computer vision method and a machine learning tobacco disease identification method. The patent is used for identifying one crop of tobacco, and the invention is mainly used for identifying multiple diseases of multiple crops.

(6) Patent document CN110188824A (application number: cn201910466618.x) discloses a small sample crop disease identification method and system. The method comprises the steps of generating an anti-network based on improved depth convolution to expand an original disease image (a first sample set) and generate the image (a second sample set) last time, verifying the second sample set, and constructing a training set by using all or part of the second sample set which passes verification and an original image which does not contain diseases; and then training the convolutional neural network based on a training set to obtain a classification model, and finally identifying. The difference between the invention and the method is mainly as follows: a) the patent document mainly aims at the identification method of small sample sets, and the invention trains all sample sets, and utilizes the DCGAN network to supplement data when the sample distribution is not balanced. b) The patent document uses the existing model to train data, and the invention proposes a new RDN classification model to train. c) The patent document mainly performs training verification on sample data of citrus canker scab, but the invention performs training on data sets of 33 diseases of 10 crops, and can obtain higher identification accuracy if other disease data sets exist.

Problems to be solved by the invention

Aiming at the problems existing in the current research situation at home and abroad, the invention mainly solves the following problems:

(1) the problem of single sample collection source is solved. The invention has the data set disclosed on the network, and also adds the data collected artificially in the natural environment and the data generated by DCGAN. The multi-source data are utilized for training, so that the over-fitting condition in the training process is avoided, and the generalization capability of the model is further improved.

(2) Aiming at the problem that the data distribution condition is not considered, the data distribution is checked and the data cleaning work is carried out by adopting a t-SNE data visualization tool.

(3) Aiming at the problem of training data imbalance, a DCGAN class equalization method is provided to ensure that training data of each class are consistent as much as possible.

(4) Aiming at the problems of low innovation degree and low recognition accuracy of a training model under various crop seed categories, the invention provides a crop disease recognition model Based on a deep Residual Dense Connection Based Network (deep Residual Dense Connection Based Network), and the recognition accuracy is further improved.

Disclosure of Invention

The method for generating the confrontation network by using the mixed data set has the advantages that the method for mixing the data sets is used, the data sets are not only disclosed, but also can be automatically acquired, and the data sets generated by the confrontation network are generated by using deep convolution. In the implementation process, firstly, the data visualization technology is used for checking the distribution of the training data, if each group of data has the condition of unbalanced distribution, the DCGAN technology can be used for generating partial images, and the training data are expanded, so that the generalization capability of the model is increased, and the identification accuracy is improved. Then a deep residual error dense connection network model is designed for crop disease identification, and 97% accuracy is achieved on a large-scale data set. The main inventive content of the invention is as follows:

(1) and a data acquisition part. In order to maximally improve the generalization capability of the model, the collected data of the public data set, such as AI CHALLENGER and Plantvillage data, also comprises data artificially collected in a natural environment.

(2) And carrying out visual analysis on the training data by utilizing a t-SNE technology to obtain the distribution condition of the data. The visual analysis is mainly divided into two steps. a) And (3) carrying out visual analysis on each type of training data by using a t-SNE technology, wherein the t-SNE is an unsupervised learning method, and the data clustering condition can be found through the visual analysis, so that the interference of junk data is avoided. b) And secondly, carrying out visual analysis on the category of the training data and the sample number of the category by using a t-SNE technology, checking the distribution condition of the data category, and carrying out related intervention in the next step if the data category is unbalanced.

(3) The unbalanced condition of the data classes can lead to the weakening of the generalization capability of the model, thereby reducing the identification accuracy. The invention provides a training sample equalization scheme based on DCGAN. Namely, the DCGAN technology is utilized to perform data enhancement on the classes with less samples, and the classes with less samples are respectively input into the DCGAN for training to generate a certain number of samples. According to the method, no special description is made on the size of the data volume, and according to experimental verification, the generalization capability of the model is improved under the condition that the sample number distribution is relatively even. After the sample is generated, the original image and the generated image are classified into one class, the t-SNE technology is utilized to carry out visual analysis, the data clustering condition is checked, and the generated sample and the original image are integrated into a whole and can be regarded as the same class.

(4) A crop disease recognition model based on RDN is provided, a Deep Residual error Network (Deep Residual Network) and a Deep Dense connection Network (Deep Dense Network) are analyzed, a Residual error connection is added to the Deep Residual error Network on the basis of an original CNN, and the problem of gradient disappearance or gradient explosion in the training process is solved; the deep Dense Connection Network takes all the outputs of the previous layer as the inputs of the next layer, so that the problem of low training efficiency is solved. Experiments show that the recognition accuracy of the model is greatly improved.

The main inventor of the invention firstly improves the Tomato crop Network, adds a pooling layer and a full-connection classification layer, and uses the pooling layer and the full-connection classification layer for image classification work, which is embodied in a paper (C.Zhou, et al, "Tomato Leaf Disease Identification by means of structured Deep crop depth Network," IEEE Access "), but the paper only researches Tomato Leaf Disease Identification, does not have related works such as a DCGAN image enhancement technology, a t-SNE data visualization technology and system development, and the Identification accuracy of only one Tomato crop is 95%. According to the invention, the model structure is further optimized, please refer to fig. 4, and methods such as DCGAN and t-SNE data operation are added, so that the model learning ability and the identification accuracy rate are enhanced. In an example experiment, the identification accuracy rate of 33 diseases of 10 crops can reach 97%.

Drawings

FIG. 1: deep residual error connection network schematic diagram

FIG. 2: deep dense connection network schematic diagram

FIG. 3: (C.Zhou, et al.2021) RDN structural diagram in paper

FIG. 4: model architecture schematic of the invention

FIG. 5: partial original image collected by the embodiment of the invention

FIG. 6: according to the embodiment of the invention, partial images of the same category are generated by using DCGAN (digital content analysis) on the basis of the acquired original image

FIG. 7: mixing the original image and the generated image together and carrying out clustering analysis by using a t-SNE visualization method

FIG. 8: originally acquired image class distribution

FIG. 9: originally acquired image and DCGAN generated image class distribution

FIG. 10: accuracy of model in training set and validation set

FIG. 11: example of crop disease recognition System

FIG. 12: data flow diagram of an embodiment of the invention

Detailed Description

For better understanding of the objects, aspects and advantages of the present invention, reference is made to the following description taken in conjunction with the accompanying drawings and the detailed description of the invention, and other advantages and effects of the invention will become apparent to those skilled in the art from the description herein, but it is not intended to limit the invention in any way. It should be noted that variations and modifications can be made by persons skilled in the art without departing from the spirit of the invention. All falling within the scope of the present invention. Some embodiments of specific examples of the present invention are described in detail below with reference to the accompanying drawings. Without conflict, the embodiments and features of the embodiments described below can be extended to all plant disease identification.

The crop disease identification method and system based on the combination of DCGAN and RDN provided by the invention comprise the following steps:

step 1: and (5) data acquisition. The multivariate and richness of the data acquisition condition can avoid overfitting of the model, the generalization capability of the model can be improved, and the accuracy of identification is directly influenced by the quality of training data. The data of the present invention are derived from the following three aspects: a) data sets published on the web, such as AI CHALLENGER and Plant village, among others; b) the data collected manually, the data collected manually are mainly collected under the natural environment, the background is relatively complicated, but have higher data quality, a) and b) collect 35538 images together in the example of the invention; c) data generated by using DCGAN are mainly used for generating a countermeasure network and performing model training by using original image data, extracting strong robustness characteristics of the training data, and then generating images of the same category by using the characteristics, wherein a part of the original images refer to fig. 5, and a part of the images generated by using DCGAN refer to fig. 6. The original image and the generated image are mixed together for clustering analysis by using a t-SNE visualization experiment, referring to FIG. 7, the experiment shows that the generated image and the original image have high fusion.

Step 2: and (5) data preprocessing work. a) And (2) data cleaning, namely performing visualization operation on each type of data by using clustering tools such as t-SNE (single-event noise analysis) and the like, wherein the data of the same type should show stronger clustering characteristics, analyzing noise data (namely data not belonging to the type) doped in each type of training data sample through the visualization operation, and manually removing the noise data so as to improve the learning capability of the model. b) And (3) data expansion, wherein the samples of each type of training data are subjected to visualization analysis by using data visualization tools such as t-SNE and the like, whether the quantity distribution of the respective types of data samples is balanced is analyzed, if the quantity difference of the samples of different types is large, the example graph 8 is referred to, the number of some type images is 4000 and more, and the number of some type images is less than 100, so that the condition that the type distribution is unbalanced occurs, and if data intervention is not performed, the type identification accuracy rate is reduced due to the fact that the types are few, so that the identification capability of the whole model is influenced. The DCGAN technology has strong feature expression capability, and the generated image retains the features of the image of the type as much as possible, so that the generated image has better fusion capability with the original image (please refer to fig. 7). Specifically, in the present example, a total of 79701 images generated by DCGAN and the original image are added, and please refer to fig. 9 for the visualization situation of each category example.

And step 3: the DCGAN generates the specific implementation steps of the countermeasure network. Specific parameters

Specifically, the DCGAN is composed of two parts, namely an arbiter and a generator: the arbiter uses LReLU for all convolutional layers and sigmod for the last layer as output. Lreul activation function role: the LReLU is used to avoid the ReLU, so that the phenomenon of over-high scoring of the image generated by the generator is prevented, the disappearance of the gradient can be relieved, and the processing capacity of the discriminator on the image characteristics is enhanced; sigmod activates the function action: converting the final result into a [0, 1] probability value as the probability that the picture is not generated by the generator;

the DCGAN generator consists of several deconvolution layers, all of which use ReLU, except for the activation function of the last layer, which uses tanh. tan h activation function role: because the last layer of the generator is the output image, and the pixel value that the ReLU activation function may output is very large, whereas the output of tanh is between [ -1,1], only simple four arithmetic operations need to be performed on the output to quickly make the output between 0 and 255, so as to output the image in the last layer; the role of the ReLU activation function is: a) adding a nonlinear function into the network to increase the learning capability of the network on the characteristics; b) the gradient disappearance condition in the back propagation process is effectively relieved; c) the dependency relationship between layers is reduced, and the capability of preventing overfitting is achieved to a certain degree.

And 4, step 4: and dividing a training set, a verification set and a test set. The invention divides training data into three parts: training set, validation set and test set. The training set and the verification set are used in the model training process, each epoch outputs a training accuracy rate, a verification accuracy rate, a training loss and a verification loss, and the model is properly adjusted and learned according to the accuracy rate and loss function data, so that the learning capability of the model is gradually improved. The test set is unknown to the model and is used to evaluate the trained model. The invention divides the data of the training set, the verification set and the test set according to the proportion of 60 percent, 20 percent and 20 percent.

And 5: and constructing the RDN recognition model. The invention combines a Deep Residual error Network model (Deep Residual Network) and a Dense connection Network model (Deep sense Network) to construct a Residual error Dense connection Network (Deep Residual Network) model, the model integrates the advantages of the Residual error connection Network and the Dense connection Network, experiments show that the model is respectively superior to the Deep Residual error Network and the Dense connection Network in accuracy, please refer to fig. 4.

Step 6: setting RDN model parameters, and carrying out model training. Specific model parameters are shown in the following table.

Specifically, the size of the input image is normalized to be an image with a uniform size, and the training batch is the number of images which can be processed by the model each time; in an optimization layer, an adapelta optimization function is utilized to minimize a loss function, and meanwhile, the learning rate can be adaptively adjusted, and the initial learning rate is set to be 0.0001; in a Residual detect module, a ReLU activation function is adopted after each layer of convolution, the normalized tensor is solved by using a LeakyReLU function, and the advantage of using the LeakyReLU is as follows: in the back propagation process, for the part of the LeakyReLU activation function input less than zero, the gradient can also be calculated, thereby solving the neuron 'death' (dying ReLU proplem) problem. The loss function is an important function for measuring the gap between the calculated output and the target of the network, the invention adopts a Cross-entropy (Cross-entropy) loss function to process the multi-classification problem, and finally adopts a softmax activation function at an output layer.

And 7: and (6) performing model training according to the parameters set in the step (6), and storing the training model. Referring to fig. 10, it can be seen that, when the number of training iterations is 161, the accuracy of the verification set is as high as 97%. And after the training is finished, saving the training model as a file in h5 format.

And 8: the model is evaluated using a test set. And loading a model, and carrying out accuracy evaluation on the test set by using a model evaluation function, wherein the accuracy rate on the test set is 96.75% by combining the embodiment of the invention.

And step 9: comparisons were made on a unified test set with existing published classification models. The following table is compared with the currently popular model, and it can be seen that the present invention maintains significant advantages in both accuracy and loss values.

Step 10: and predicting the single or batch input images by using the trained model. And inputting a picture by combining the example, predicting the disease category by using the trained model, and outputting the Top-1 probability value of the crop disease.

Step 11: the trained model is applied to a crop leaf disease identification system, a foreground is used for data input and output, a background is used for calculation, and please refer to fig. 11 for a design and operation interface. (note: the WeChat applet utilized in this example is only one example, and other software programs developed using the concepts of the present invention are considered to be within the scope of the present invention). In summary, please refer to fig. 12 for a data flow chart of the present invention.

It will be appreciated by those skilled in the art that, in addition to implementing the system, apparatus and various modules thereof provided by the present invention in the form of pure computer readable program code, the system, apparatus and various modules thereof provided by the present invention can be implemented in the same form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers and the like by entirely logically programming the method steps. Therefore, the system and the modules thereof provided by the present invention can be regarded as a hardware component, and the modules for implementing various programs can also be regarded as structures in the hardware component; modules for performing various functions may also be considered to be both software programs for performing the methods and structures within hardware components.

The foregoing description of specific embodiments of the present invention has been presented. It is to be understood that the present invention is not limited to the specific embodiments described above, and that various changes or modifications may be made by one skilled in the art within the scope of the appended claims without departing from the spirit of the invention.

Claims

1. A crop disease identification method and system based on DCGAN and RDN are characterized by comprising the following steps:

s1: firstly, data acquisition is carried out;

s2: the method comprises the following steps of performing data preprocessing work, and performing data visualization, data cleaning, data generation and other technologies to guarantee the accuracy and the distribution balance of a training data set, and specifically comprises the following steps:

s21: data visualization, namely performing clustering visualization operation on each type of data by using a t-SNE data visualization tool, and analyzing the data distribution condition according to a clustering visualization result;

s22: data cleaning, namely analyzing noise data (namely data not belonging to the category) doped in each type of training data sample according to the data visualization result in the step S21, and clearing the noise data;

s23: data expansion, namely performing visual analysis on the number of samples contained in the sample classes by using a t-SNE data visualization tool, analyzing whether the number distribution of the data samples in the respective classes is balanced or not, and if the number distribution of the data samples in the respective classes is not balanced, performing data enhancement operation by using DCGAN (data enhancement algorithm) to expand a training data set;

s3: the method comprises the following steps of dividing a training set, a verification set and a test set, and dividing a data set into: training set 60%, validation set 20%, and test set 20%;

s4: the method comprises the steps of constructing a deep residual error network (RDN) identification model, fusing and optimizing two network models, removing a flat layer, a dense layer and a classifier of an intensive network, combining the RDN identification model and the dense network into a residual error intensive module, connecting the residual error intensive module and the deep residual error network again in a network global environment to form the deep residual error intensive network model, combining the advantages of the deep residual error network and the deep residual error intensive network, improving the learning capability of the model to improve the identification accuracy, and enhancing the calculation efficiency;

s5: setting model parameters, and loading a training set and a verification set for model training;

s6: and (5) performing model training according to the parameters set in the step (S5) and saving the training model.

S7: loading a model, and carrying out accuracy evaluation on the model on a test set by using a model evaluation function;

s8: and applying the trained model to a crop leaf disease identification system, wherein a foreground is used for inputting leaf images, a background is used for calculating, and finally, disease categories and probability values belonging to the categories are returned.

2. The method and system for crop disease identification based on DCGAN and RDN of claim 1, wherein said DCGAN countermeasure generates 3 channel images with network input image size 196X 196, batch 64, initial learning rate 0.0002, optimization function using adam, number of iterations 1000, loss function Cross-entry, activation function LReLU, ReLU, sigmoid, tanh.

3. The crop disease identification method and system based on DCGAN and RDN as claimed in claim 1, wherein in RDN model, 3 channel image with size 196X 196 of image is inputted, the number of filters of the first convolutional layer is changed to 128, the number of filters of the second convolutional layer is still 64, the RDB module of the present invention is added with tensor batch standardization, and batch standardization is performed after each convolution.

4. The crop disease identification method and system based on DCGAN and RDN as claimed in claim 1, wherein the tensor after the linkage of the continate is convolved twice, each convolution is followed by a batch normalization, the tensor is then connected with conv1_ pool by add (residual), the obtained tensor is sent into two convolution layers, the number of filters in the first convolution layer is 64, the number of filters in the second convolution layer is 256, kernel _ szie is 3, strings is (1,1), the batch normalization is not performed after the convolution in the first layer, the batch normalization and the maximum pooling are performed after the convolution in the second layer, and the parameters of the maximum pooling layer are pool _ size is 2 and strings is 2.

5. The crop disease identification method and system based on DCGAN and RDN as claimed in claim 1, wherein the tensor after the last residual connection is convolved once, the number of filters is 256, kernel _ size is 3, and strides is (1,1), then the tensor is flattened in the first-order tensor space by using a Flatten layer, feature extraction is performed by using three fully-connected layers, the neural units of the three fully-connected layers are 256, 128, and 64 in sequence, after the first fully-connected layer, Dropout with a parameter of 0.3 is performed once, after the second fully-connected layer, L2 regularization with a parameter of 0.0005 is performed once, and finally the classifier activation function is softmax.

6. The crop disease recognition method and system based on DCGAN and RDN as claimed in claim 1, wherein the system is divided into three parts, namely a data acquisition module, a recognition module and an output module, wherein the data acquisition module can be a photo stored by a terminal or a photo taken by a camera, and then the photo is transmitted to a remote server for recognition by using a network, the recognition module calculates by a trained network model, and finally the calculation result is output to a terminal device by using the network.