CN107016406A - The pest and disease damage image generating method of network is resisted based on production - Google Patents

Abstract

The invention relates to a method for generating images of pests and diseases based on a generative confrontation network, which solves the defect of less sampling images of pests and diseases compared with the prior art. The invention includes the following steps: collecting and preprocessing the training images; constructing a discrimination network and a generation network based on a deep convolutional neural network model; training the discrimination network and the generation network; generating images of diseases and insect pests according to the trained generation network. The invention can generate a large number of similar real images of pests and diseases based on a small number of existing images of pests and diseases, provides sample images for image recognition of pests and diseases, and solves the problem of relatively few images of pests and diseases in the field and high acquisition cost.

Description

Pest image generation method based on generative adversarial network

technical field

The invention relates to the technical field of image generation, in particular to a method for generating images of diseases and insect pests based on generative confrontation networks.

Background technique

Pests and diseases are the enemies of crops. They occur throughout the growth period of crops and can cause a large reduction in crop yield. Effective identification of pests and diseases through software technology has become a research hotspot in the industry. However, the current pest diagnosis method based on image recognition requires a large number of pest image samples as the support of raw data. However, in practical applications, there are relatively few samples of existing images of pests and diseases, and the workload of collecting images of field diseases and insect pests is relatively difficult and large.

Therefore, how to use a small number of existing pest images to realize the regeneration of pest images has become an urgent technical problem to be solved.

Contents of the invention

The purpose of the present invention is to solve the defect of few sampling images of pest images in the prior art, and provide a method for generating pest images based on generative confrontation network to solve the above problems.

In order to achieve the above object, the technical scheme of the present invention is as follows:

A method for generating images of pests and diseases based on a generative confrontation network, comprising the following steps:

Collect and preprocess the training images, collect several images as training images, and normalize the size of all training images, process them into 256×256 pixels, and obtain several training samples;

Construct a discriminant network and a generative network based on a deep convolutional neural network model;

Train the discriminative network and the generative network;

Generate images of pests and diseases according to the trained generation network.

The described construction of discriminant network and generation network based on deep convolutional neural network model comprises the following steps:

Use the deep convolutional neural network model to construct the discriminant network, set the number of network layers of the deep convolutional neural network model to 5 layers, of which the first 3 layers are convolutional layers, the fourth layer is a fully connected layer, and the last layer is an output layer. The number of nodes in the output layer is 1, its input is an image, and the size of the input image is 256*256;

Use the deep convolutional neural network model to construct the generative network, set the number of network layers of the deep convolutional neural network model to 4 layers, of which the first 3 layers are deconvolution layers, the last layer is the output layer, and the number of nodes in the output layer is 256*256, whose input is random noise.

The described training of discriminative network and generation network comprises the following steps:

Set the loss function, the formula is as follows:

Among them, D(x) is the output of the discriminant network on the training data set, x~P _data (x) is the real probability distribution of the data set, and D(G(z)) is the output of the discriminant network on the image generated by the generating network , z～P _z (x) is the probability distribution of the training data set generated by the network simulation, z is a random vector, In order to enable the discriminative network to distinguish the input real data, In order to enable the generation network to deceive the discriminative network;

Use V(D,G) as the loss function when training the discriminative network, and use it when training the generative network as a loss function;

The generation of discriminative network training data, assuming that the training batch size is 50, then 25 positive samples are randomly selected from the training samples, then the generation process of 25 negative samples is as follows:

Generate 25 random vectors;

25 random vectors are used as the input of the generation network to obtain 25 forged data, which are calibrated as negative samples of the discriminant network;

Generate network training data generation, set the training batch size to 50, then the generation process of 50 positive samples is as follows:

Generate 50 random vectors;

Take 50 random vectors as the input of the generator network, get 50 fake data, and mark it as a positive sample of the generator network;

To train the network, the specific steps are as follows:

Set the hyperparameter k, and then perform the training of the generation network after each training of the discriminant network for k times;

The discriminative network is trained,

Select m noise samples, the prior probability distribution is p _g (z), marked as {z ⁽¹⁾ ,...,z ^(m) };

Select m training samples, the probability distribution is p _data (x), marked as {x ⁽¹⁾ ,...,x ^(m) };

According to the stochastic gradient descent method, the parameters of the discriminant network are modified, and the formula for calculating the stochastic gradient is as follows:

Generate a network for training,

Select m noise samples, the prior probability distribution is p _g (z), marked as {z ⁽¹⁾ ,...,z ^(m) }, according to the stochastic gradient descent method, modify the parameters of the generation network, which calculates the random The gradient formula is as follows:

The discriminant network judges the true probability of the picture. When the discriminant network judges that the probability of the picture being a training image tends to 0.5, the training is completed.

The described generation of images of diseases and insect pests according to the trained generation network includes the following steps:

Randomly select n noise samples from the training samples, the prior probability distribution is p _g (z), marked as {z ⁽¹⁾ ,...,z ⁽ⁿ⁾ };

Input the noise samples into the trained generation network respectively, and generate n images of diseases and insect pests.

Beneficial effect

Compared with the prior art, the method for generating images of diseases and insect pests based on the generative confrontation network of the present invention can generate a large number of similar real images of diseases and insect pests based on a small number of existing images of diseases and insect pests, providing sample images for image recognition of diseases and insect pests, and solving the practical problems The problems of relatively few images of field diseases and insect pests and high acquisition costs.

Description of drawings

Fig. 1 is a method sequence diagram of the present invention.

detailed description

In order to have a further understanding and understanding of the structural features of the present invention and the achieved effects, the preferred embodiments and accompanying drawings are used for a detailed description, as follows:

As shown in Figure 1, a method for generating images of diseases and insect pests based on generative confrontation network according to the present invention comprises the following steps:

The first step is to collect and preprocess the training images, collect several images as training images, and normalize the size of all training images to 256×256 pixels to obtain several training samples.

In the second step, the discriminant network and the generation network are constructed based on the deep convolutional neural network model. The deep convolutional neural network emphasizes the depth of the model structure, highlights the importance of feature learning, and can learn the essential features of the image. Pest images are affected by various factors such as farmland background, illumination, occlusion, etc., so they are more complicated than general images. Therefore, constructing a discriminant network and a generative network based on a deep convolutional neural network model can generate more realistic images of pests and diseases after completing network training.

The specific steps are as follows:

(1) Construct a discriminative network using a deep convolutional neural network model. Set the number of network layers of the deep convolutional neural network model to 5 layers, of which the first 3 layers are convolutional layers, the fourth layer is a fully connected layer, the last layer is an output layer, the number of nodes in the output layer is 1, and its input is An image, the input image size is 256*256.

The more classic deep convolutional neural network model is LeNet-5. This model has a better effect on handwritten character recognition. It has only two convolutional layers, but it does not work well when applied to more complex images of diseases and insect pests. Through experiments, it is found that the effect is better when the convolutional layer is increased to 3 layers. Therefore, the number of roll bases is set to 3 here, and the fully connected layer and output layer must exist, so a total of 5 layers are designed here. (The expression of the number of layers here does not include the input layer in the usual way of expression)

(2) Use the deep convolutional neural network model to construct the generative network, set the network layer number of the deep convolutional neural network model to 4 layers, of which the first 3 layers are deconvolution layers, the last layer is the output layer, and the nodes of the output layer The number is 256*256, and its input is random noise. Random noise is an artificial disturbance, and general programming languages provide functions to create random noise.

Unlike the discriminative network, the generative network model does not have a convolutional layer and a fully connected layer, only a deconvolutional layer and an output layer.

The third step is to train the discriminative network and the generative network. The specific steps are as follows:

(1) Set the loss function, the formula is as follows:

Among them, D(x) is the output of the discriminant network on the training data set, x~P _data (x) is the real probability distribution of the data set, and D(G(z)) is the output of the discriminant network on the image generated by the generating network , z～P _z (x) is the probability distribution of the training data set generated by the network simulation, z is a random vector, In order to enable the discriminative network to distinguish the input real data, In order to enable the generative network to fool the discriminative network.

Use V(D,G) as the loss function when training the discriminative network, and use it when training the generative network as a loss function.

(2) The generation of the discriminative network training data, assuming that the training batch size is 50, then 25 positive samples are randomly selected from the training samples, then the generation process of 25 negative samples is as follows:

A. Generate 25 random vectors;

B. Take 25 random vectors as the input of the generator network, get 25 forged data, and mark it as the negative sample of the discriminant network.

(3) Generate the generation of network training data, assuming that the training batch size is 50, then the generation process of 50 positive samples is as follows:

A. Generate 50 random vectors;

B. Take 50 random vectors as the input of the generating network, get 50 fake data, and mark them as positive samples of the generating network.

(4) training network, its specific steps are as follows:

A. Set the hyperparameter k, and perform the training of the generation network after every k times of discriminant network training. That is, the training of the generation network is carried out after k times of discriminant network, the training method of the discriminant network is as described in the step of training the discriminative network, and the training method of the generative network is as described in the step of training the generative network.

B. The discriminative network is trained.

Select m noise samples, the prior probability distribution is p _g (z), marked as {z ⁽¹⁾ ,...,z ^(m) };

Select m training samples, the probability distribution is p _data (x), marked as {x ⁽¹⁾ ,...,x ^(m) };

According to the stochastic gradient descent method, the parameters of the discriminant network are modified, and the formula for calculating the stochastic gradient is as follows:

Represents the gradient, θ represents the network parameters, and θ _d represents the parameters of the discriminant network.

C. Generate a network for training,

Select m noise samples, the prior probability distribution is p _g (z), marked as {z ⁽¹⁾ ,...,z ^(m) }, according to the stochastic gradient descent method, modify the parameters of the generation network, which calculates the random The gradient formula is as follows:

Represents the gradient, θ represents the network parameters, and θ _g represents the parameters of the generated network.

D. The discriminant network judges the true probability of the picture. When the discriminant network judges that the probability that the picture is a training image tends to 0.5, the training is completed.

The fourth step is to generate images of diseases and insect pests according to the trained generation network, input noise through the generation network, and output pictures. The specific steps are as follows:

(1) Randomly select n noise samples from the training samples, the prior probability distribution is p _g (z), marked as {z ⁽¹⁾ ,...,z ⁽ⁿ⁾ }.

(2) Input the noise samples into the trained generation network to generate n pest images.

The basic principles, main features and advantages of the present invention have been shown and described above. Those skilled in the art should understand that the present invention is not limited by the above-mentioned embodiments. What are described in the above-mentioned embodiments and the description are only the principles of the present invention. Variations and improvements, which fall within the scope of the claimed invention. The scope of protection required by the present invention is defined by the appended claims and their equivalents.

Claims (4)

1. A method for generating images of diseases and insect pests based on generative confrontation network, characterized in that, comprising the following steps: 11) Collect and preprocess the training images, collect several images as training images, and normalize the size of all training images, process them into 256×256 pixels, and obtain several training samples; 12) Construct a discriminant network and a generated network based on a deep convolutional neural network model; 13) Train the discrimination network and the generation network; 14) Generate images of diseases and insect pests according to the trained generation network. 2. the method for generating images of diseases and insect pests based on generative confrontation network according to claim 1, is characterized in that, constructing discriminant network and generating network based on deep convolutional neural network model comprises the following steps: 21) Use a deep convolutional neural network model to construct a discriminant network, set the number of network layers of the deep convolutional neural network model to 5 layers, of which the first 3 layers are convolutional layers, the fourth layer is a fully connected layer, and the last layer is output layer, the number of nodes in the output layer is 1, its input is an image, and the size of the input image is 256*256; 22) Use a deep convolutional neural network model to construct a generative network, set the number of network layers of the deep convolutional neural network model to 4 layers, of which the first 3 layers are deconvolution layers, the last layer is the output layer, and the number of nodes in the output layer is The number is 256*256, and its input is random noise. 3. the method for generating images of diseases and insect pests based on generative confrontation network according to claim 1, characterized in that, the described training of discriminant network and generation network comprises the following steps: 31) Set the loss function, the formula is as follows: Among them, D(x) is the output of the discriminant network on the training data set, x~P _data (x) is the real probability distribution of the data set, and D(G(z)) is the output of the discriminant network on the image generated by the generating network , z～P _z (x) is the probability distribution of the training data set generated by the network simulation, z is a random vector, In order to enable the discriminative network to distinguish the input real data, In order to enable the generation network to deceive the discriminative network; Use V(D,G) as the loss function when training the discriminative network, and use it when training the generative network as a loss function; 32) Discriminate the generation of network training data, assuming that the training batch size is 50, then 25 positive samples are randomly selected from the training samples, then the generation process of 25 negative samples is as follows: 321) generate 25 random vectors; 322) 25 random vectors are used as the input of the generation network to obtain 25 forged data, and are marked as negative samples of the discriminant network; 33) Generate the generation of network training data, assuming that the training batch size is 50, then the generation process of 50 positive samples is as follows: 331) generate 50 random vectors; 332) 50 random vectors are used as the input of the generating network, and 50 forged data are obtained, and marked as positive samples of the generating network; 34) training network, its specific steps are as follows: 341) Set the hyperparameter k, and then perform the training of the generation network after every k times of discriminant network training; 342) discriminant network for training, Select m noise samples, the prior probability distribution is p _g (z), marked as {z ⁽¹⁾ ,...,z ^(m) }; Select m training samples, the probability distribution is p _data (x), marked as {x ⁽¹⁾ ,...,x ^(m) }; According to the stochastic gradient descent method, the parameters of the discriminant network are modified, and the formula for calculating the stochastic gradient is as follows: 343) generating a network for training, Select m noise samples, the prior probability distribution is p _g (z), marked as {z ⁽¹⁾ ,...,z ^(m) }, according to the stochastic gradient descent method, modify the parameters of the generation network, which calculates the random The gradient formula is as follows: 344) The discriminant network judges the true probability of the picture, and when the discriminant network judges that the probability that the picture is a training image tends to 0.5, the training is completed. 4. the method for generating images of diseases and insect pests based on generative confrontation network according to claim 1, characterized in that, generating the images of diseases and insect pests according to the trained generation network comprises the following steps: 41) Randomly select n noise samples from the training samples, the prior probability distribution is p _g (z), marked as {z ⁽¹⁾ ,...,z ⁽ⁿ⁾ }; 42) Input the noise samples into the trained generation network respectively, and generate n images of diseases and insect pests.