KR20210047230A - Fruit tree disease Classification System AND METHOD Using Generative Adversarial Networks - Google Patents

Abstract

According to one embodiment of the present invention, provided is a fruit tree disease and insect damage classification system using a generative adversarial network (GAN), which comprises: a data processing unit including a data pre-processing portion which pre-processes data of an original image including a disease and insect damage to reduce the data operation amount and increase a data operation speed for reducing resources used during data learning, and a post-processing portion which stores and filters result data of the pre-processing; and a network unit including a GAN network which generates a similar image corresponding to the result data of the pre-processing through the GAN, and a classification network which uses the generated similar image to train a classification model, and determines and classifies the type of the disease and insect damage of an input image based on the trained classification model.

Description

Fruit tree disease classification system and method using generative adversarial networks

Embodiments of the present invention relate to a system and method for classifying fruit diseases and pests, and in more detail, a generative adversarial neural network capable of generating data that can be used for learning using Generative Adversarial Networks (GAN). It relates to a system and method for classification of fruit tree diseases and pests using

The smart farm is evolving more recently with the application of the 4th industrial revolution technology. Automation, intelligence, and connection by incorporating new technologies such as information and communication technology (ICT), Internet of Things (IoT), big data, cloud, and artificial intelligence (AI) with the growth and environment of crops and livestock. It enables remote control as well as services such as etc. As a result, it is possible to perform various necessary tasks by itself, including environmental control operations required according to the current state of the farm.

The need for smart farms is deeply related to the food shortage caused by global climate change and population growth. The demand for food is increasing due to the increase of the world population, but the number of people to farm due to the decline of the agricultural population and an aging population is becoming increasingly scarce. In other words, the population increases, the area of cultivation of crops decreases due to urbanization, and the farmers at the production sites are aging. Even in Korea, the average age of top farmers in 2017 is 67, and this average age is only increasing with time. For this reason, it can be seen that the expansion of smart farms is inevitable.

However, smart farms are concentrated in the smart horticultural field, and in recent years their application fields have spread to smart horticulture, smart livestock, smart fields, and smart distribution, but these are also used only in a limited space called facilities.

The reason why you have to work only within the facility is to control physical devices to help the growth of crops, but also to collect data necessary for learning. In order to maximize the growth potential of a crop, it is necessary to collect the data of the crop, because there are many problems in collecting the data of all crops.

In addition, the problem between the smart farm and the data is that the data on diseases that can occur in crops are different for each crop, and there are diseases that occur only in specific crops, and it is difficult to obtain sufficient data to use for learning, and even the data provided by class It is said that it adversely affects learning due to data imbalance.

As a related prior art, there is Korean Patent Publication No. 10-1933657 (title of the invention: a method and apparatus for detecting and diagnosing crop diseases and pests using deep learning, registration date: 2018.12.21.).

An embodiment of the present invention amplifies the amount of data used for learning and generates data of various classes by generating data similar to actual data by learning data through a generative adversarial neural network and receiving used data as input values. Thus, a system and method for classifying fruit tree diseases and pests using a generative adversarial neural network capable of resolving the data imbalance for each class are provided.

The problem to be solved by the present invention is not limited to the problem(s) mentioned above, and another problem(s) not mentioned will be clearly understood by those skilled in the art from the following description.

The fruit tree disease classification system using a generative adversarial neural network according to an embodiment of the present invention is calculated by reducing the amount of data computation by preprocessing the data of the original image including the disease and pests in order to reduce the resources used when learning data. A data processing unit including a data preprocessing unit for speeding up a data preprocessing unit and a data post processing unit for storing or filtering result data of the preprocessing; And a GAN network that generates a similar image corresponding to the result data of the preprocessing through a generative hostile neural network (GAN), and a classification model using the generated similar image, and input based on the learned classification model. It includes a network unit including a classification network that determines and classifies the types of diseases and pests of the image.

The data preprocessing unit downsizing by adjusting the size of the original image; A region of interest extracting unit for extracting a region of interest from the downsized original image; And an image contrast processor configured to contrast the extracted image of the ROI.

The ROI extractor may extract the ROI based on an RGB average value calculated using a sobel mask.

The data post-processing unit stores the result data of the pre-processing, and when the generated similarity images are collected in a predetermined amount or more, the data storage unit sends to the classification network; And a data filter configured to filter the result data of the pre-processing and filter data having a quality lower than a predetermined reference value among the generated similar images.

The GAN network may generate the similar image through the generative adversarial neural network based on a sample image that does not contain a disease or pest.

The GAN network repeats the process of generating the similarity image until it reaches a preset predetermined number, and the classification network uses the predetermined number of similar images when the similarity image reaches a predetermined number. By learning, it is possible to increase the amount of data used for learning and to solve the data imbalance for each class.

The generative adversarial neural network is designed identically except for the existing DCGAN and the discriminator, and the discriminator takes into account that the original image has a certain shape of a leaf and is the subject of a disease occurring in the leaf. In DCGAN, the size of the convolution layer is modified from 5*5 to 3*3 to reduce the amount of computation and save memory, and the activation function will be designed to generate similar images of various classes using Sigmoid instead of Relu. I can.

In order to reduce the resources used when learning data, the method of classifying fruit trees diseases and pests using a generative adversarial neural network according to an embodiment of the present invention is calculated by reducing the amount of data computation by preprocessing the data of the original image including the disease Speeding up; Storing or filtering result data of the pre-processing; Generating a similar image corresponding to the result data of the preprocessing through a generative hostile neural network (GAN); Learning a classification model using the generated similarity image; And determining and classifying the type of disease and pest of the input image based on the learned classification model.

The step of preprocessing the data of the original image to reduce the amount of data computation to speed up the calculation speed includes: downsizing by adjusting the size of the original image; Extracting a region of interest from the downsized original image; And performing a contrast image on the extracted image of the ROI.

Generating the similarity image may include generating the similarity image through the generative adversarial neural network based on a sample image that does not contain pests.

The generative adversarial neural network is designed identically except for the existing DCGAN and the discriminator, and the discriminator takes into account that the original image has a certain shape of a leaf and is the subject of a disease occurring in the leaf. In DCGAN, the size of the convolution layer is modified from 5*5 to 3*3 to reduce the amount of computation and save memory, and the activation function will be designed to generate similar images of various classes using Sigmoid instead of Relu. I can.

Details of other embodiments are included in the detailed description and the accompanying drawings.

According to an embodiment of the present invention, a smart farm can be applied to various crops by amplifying a small amount of data and generating data of various classes, and a higher hit rate can be realized than the existing method.

According to an embodiment of the present invention, there is an advantage that not only can generate a sufficient amount of data for learning by amplifying a small amount of data, but also apply it to an existing system and utilize it. It can compensate for the problem of limiting the farm.

According to an embodiment of the present invention, it is possible to learn a crop that has the same disease to generate disease data of the crop, amplify a small amount of data to generate a sufficient amount of data for learning. You can use it by applying it to the system as it is.

FIG. 1 is a block diagram illustrating a system for classifying fruit tree diseases and pests using a generative hostile neural network according to an embodiment of the present invention.
FIG. 2 is a block diagram illustrating a detailed configuration of a data preprocessor of FIG. 1.
3 is a diagram illustrating an image matrix that appears when a color image is loaded to extract a region of interest.
4 and 5 are flowcharts illustrating a method for classifying fruit tree diseases and pests using a generative hostile neural network according to an embodiment of the present invention.
6 is a diagram showing a structure of a discriminator model of a generative adversarial neural network applied to an embodiment of the present invention.
7 is a diagram illustrating a result of implementing an ROI.
8 is an exemplary view showing an actual black rot image.
9 is an exemplary view showing an image of an original blueberry leaf and a blueberry leaf with Black rot generated based on it.
10 is a table showing classification accuracy of a system according to an embodiment of the present invention compared with an existing system.

Advantages and/or features of the present invention, and a method of achieving them will become apparent with reference to the embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below, but will be implemented in a variety of different forms, only the present embodiments are intended to complete the disclosure of the present invention, and common knowledge in the technical field to which the present invention pertains. It is provided to completely inform the scope of the invention to those who have, and the invention is only defined by the scope of the claims. The same reference numerals refer to the same elements throughout the specification.

In addition, a preferred embodiment of the present invention to be implemented below is already provided in each system function configuration in order to efficiently describe the technical components constituting the present invention, or system functions commonly provided in the technical field to which the present invention belongs. The configuration will be omitted as much as possible, and a functional configuration that should be additionally provided for the present invention will be mainly described. If a person of ordinary skill in the art to which the present invention pertains will be able to easily understand the functions of the components previously used among the functional configurations that are not shown below and are omitted, the configurations omitted as described above. The relationship between the elements and the constituent elements added for the present invention will also be clearly understood.

In addition, in the following description, "transmission", "communication", "transmission", "receive" of signals or information and other terms with similar meanings refer to direct transmission of signals or information from one component to another. Not only that, but it includes things that are transmitted through other components. In particular, "transmitting" or "transmitting" a signal or information to a component indicates the final destination of the signal or information and does not mean a direct destination. The same is true for "reception" of signals or information.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

FIG. 1 is a block diagram illustrating a system for classifying fruit tree diseases and pests using a generative adversarial neural network according to an embodiment of the present invention.

Referring to FIG. 1, a system for classifying fruit trees diseases and pests using a generative hostile neural network according to an embodiment of the present invention includes a data processing unit 120 and a network unit 110. Can be.

The data processing unit 120 may include a data preprocessing unit 122 and a data post processing unit 124 to manage data preprocessing and postprocessing.

The data preprocessor 122 may perform a function of speeding up the computation speed by reducing the amount of data computation by preprocessing data of an original image including diseases and pests in order to reduce resources used when learning data.

To this end, the data preprocessing unit 122 may include an image size adjusting unit 210, an ROI extracting unit 220, and an image contrast processing unit 230 as shown in FIG. 2. For reference, FIG. 2 is a block diagram illustrating a detailed configuration of the data preprocessor 122 of FIG. 1.

The image size adjusting unit 210 may downsizing by adjusting the size of the original image including diseases and pests. Accordingly, the image size adjusting unit 210 may reduce the size of the original image, thereby reducing the memory space occupied by the system and reducing the amount of computation.

The region of interest extracting unit 220 may extract a region of interest (RoI) from the downsized original image. To this end, the ROI extractor 220 may extract the ROI from the downsized original image based on an RGB average value calculated using a sobel mask. Accordingly, the region of interest extracting unit 220 may reduce the amount of computation by reducing data other than the region of interest.

The region of interest literally means the region of interest in the image. When an object is detected or detected by processing an image, the detected area may be referred to as an ROI. There are three main reasons for designating a region of interest. The first is to remove unnecessary images in the area around the object, and the second is to reduce the amount of computation and resources by removing them in this way. Finally, it is because it can increase the accuracy of learning. By removing unnecessary information in advance, since only the necessary part is used for learning, accuracy can be improved through this.

In the case of a color image, each pixel has R, G, and B values. In other words, it is expressed as a multi-channel image that has three pieces of information in one pixel. On the other hand, in the case of a black-and-white image, since each pixel has one information, it is called a single-channel.

3 is a diagram illustrating an image matrix that appears when a color image is loaded to extract a region of interest. 3, the matrix on the left is a color image. In general, it looks two-dimensional, but it is actually composed of three single channels as shown on the right side of the drawing. The value of the corresponding channel in the image varies depending on the original image, and by designating a range for each channel and redrawing only the pixels within it, only areas with a specific color can be specified and separated.

The image contrast processing unit 230 may contrast the extracted image of the ROI. Accordingly, the image contrast processing unit 230 may play a role of remarkably adjusting the characteristics of data.

The data post-processing unit 124 may perform a function of storing or filtering result data of the pre-processing by the data pre-processing unit 122. To this end, the data post-processing unit 124 may include a data storage unit and a data filter unit, although not shown in the drawing.

The data storage unit stores the result data of the pre-processing by the data pre-processing unit 122, and when the similar images generated by the network unit 110 to be described later are collected in a predetermined amount or more, it is classified by the network unit 110 Sending to the network 114 may be performed.

The data filter unit may filter result data of preprocessing by the data preprocessor 122. In addition, the data filter unit may filter data having a quality lower than a predetermined reference value among the generated similar images. The filtered data may be transferred to and stored in the data storage unit.

The network unit 110 may include a GAN network 112 and a classification network 114.

The GAN network 112 may generate a similar image corresponding to the result data of preprocessing by the data preprocessor 122 through a generative hostile neural network (GAN).

In this case, the GAN network 112 may generate the similar image through the generative adversarial neural network based on a sample image that does not contain a disease or a pest. Through this, the GAN network 112 may generate an image similar to a diseased image from the original image. In addition, the GAN network 112 may repeat the process of generating the similar image until it reaches a predetermined number.

The classification network 114 may learn a classification model by using the similar image generated by the GAN network 112, and classify the input image by determining the type of pests based on the learned classification model.

When the number of similar images generated by the GAN network 112 reaches a certain number, the classification network 114 learns the classification model using the certain number of similar images, thereby determining the amount of data used for learning. It can increase and resolve the data imbalance by class (type of disease and pest).

A generative adversarial neural network (GAN) is a neural network structure composed of two networks. One network confronts the other network and learns. Equation 1 below represents an equation of a generative adversarial neural network.

[Equation 1]

In Equation 1, x~P _data (x) denotes an image generated by a probability distribution for an actual image, and z~P _z (z) denotes an image generated using noise. D(x) represents the discriminator, and a value between 0 and 1 indicating the probability that the image is real is displayed. In the case of D(G(z)), the image created by the creator is classified through a discriminator, which also has a value between 0 and 1, which means the probability that the image is real.

In order to maximize Equation 1, the value of D(G(z)) should be close to 1 and the value of D(x) should be close to 1. In this way, the learning method of the generative adversarial neural network is that the generator and the discriminator solve the Minmax problem and learn.

In one embodiment of the present invention, the generative adversarial neural network may be designed identically except for the existing DCGAN (Deep Convolutional) and the discriminator (Discriminator). For reference, the DCGAN refers to a network in the form of a combination of unsupervised learning and a convolutional neural network (CNN).

At this time, the discriminator is the size of the convolution layer in the conventional DCGAN as shown in FIG. 6, considering that the original image has a certain shape of a leaf and is the subject of a disease occurring in the leaf. It can be designed to reduce the amount of computation and save memory by modifying from 5*5 to 3*3, and to generate similar images of various classes using Sigmoid instead of Relu as the activation function.

The system described above may be implemented as a hardware component, a software component, and/or a combination of a hardware component and a software component. For example, the devices and components described in the embodiments include, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable array (FPA), It may be implemented using one or more general purpose or special purpose computers, such as a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications executed on the operating system. Further, the processing device may access, store, manipulate, process, and generate data in response to the execution of software. For the convenience of understanding, although it is sometimes described that one processing device is used, one of ordinary skill in the art, the processing device is a plurality of processing elements and/or a plurality of types of processing elements. It can be seen that it may include. For example, the processing device may include a plurality of processors or one processor and one controller. In addition, other processing configurations are possible, such as a parallel processor.

The software may include a computer program, code, instructions, or a combination of one or more of these, configuring the processing unit to operate as desired or processed independently or collectively. You can command the device. Software and/or data may be interpreted by a processing device or, to provide instructions or data to a processing device, of any type of machine, component, physical device, virtual equipment, computer storage medium or device. , Or may be permanently or temporarily embodyed in a transmitted signal wave. The software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

4 is a flow chart illustrating a method for classifying fruit tree diseases and pests using a generative adversarial neural network according to an embodiment of the present invention.

The fruit tree disease and pest classification method described herein is only one embodiment of the present invention, and various steps may be added as follows as needed, and the following steps may also be performed by changing the order, so the present invention It is not limited to each step and its sequence described below.

First, referring to FIG. 4, in step 410, the fruit tree disease and pest classification system 100 may downsize by adjusting the size of the original image including the disease and pest (Image Resize).

Next, in step 420, the fruit tree disease and pest classification system 100 may extract a region of interest (RoI) from the downsized original image (Crop Region of interest).

Next, in step 430, the fruit tree disease and pest classification system 100 may contrast the extracted image of the ROI (Contrast Image).

Next, in step 440, the fruit tree disease and pest classification system 100 may generate a similar image corresponding to data as a result of pre-processing (410, 420, 430) as described above through a generative hostile neural network (GAN). (Generate Image using GAN).

For reference, the generative adversarial neural network designed in an embodiment of the present invention may design a model based on DCGAN in an existing neural network. The generator is the same as the existing DCGAN, but the discriminator can be transformed to fit the prepared data. 6 is a diagram illustrating a structure of a discriminator model of a generative adversarial neural network applied to an embodiment of the present invention.

Referring to FIG. 6, since the image given as an input has a certain shape of a leaf and the disease occurring in the leaf is the subject, the discriminator can correct it accordingly. First, the size of the convolution layer can be modified from 5x5 to 3x3 to reduce the amount of computation and to save memory, and the activation function can be designed to generate images of various classes using Sigmoid instead of Relu.

Next, in step 450, the fruit tree disease and pest classification system 100 may filter and output the generated similarity image (Filter out).

Next, in step 460, it may be determined whether the quality of the generated similarity image exceeds a predetermined reference value (Generated Data is above the standard).

In this case, if the reference value is not exceeded (in the “no” direction of 460), the operation may be returned to step 440. On the other hand, if the reference value is exceeded (in the "yes" direction of 460), in step 470, the fruit tree disease and pest classification system 100 learns a classification model using the generated similarity image, and the learned classification model Based on the data classification, it is possible to classify data by determining the type of disease and pest of the input image (Data Classification).

5 is a flowchart illustrating a process of extracting an ROI by partially adopting a Sobel mask method based on RGB extraction.

Referring to FIG. 5, first, in step 510, an image is searched (Navigate Pixel). Thereafter, an average value is calculated in step 520 (Average calculation). Here, the average value is calculated using a Sobel mask.

Next, after calculating the average value in step 530, the inspection is started for each pixel (Check the 8th direction of the pixel).

If it does not exceed the previously obtained average value (in the "no" direction of 540), the next pixel is examined in step 550 (Navigate Diagonal Pixel).

However, if it exceeds the previously obtained average value (in the "yes" direction of 540), it is checked in step 560 whether the pixel is stored (check if stored coordinate).

At this time, if the pixel is not stored (in the "no" direction of 560), the pixel is stored in step 570 and moves to the pixel to perform inspection (save coordinate and move that pixel). At this time, the inspection direction of the pixel proceeds constantly.

On the other hand, if the pixel is stored (in the "yes" direction of 560), the search is terminated in step 580 and a mask is created with the stored pixels (Create mask).

Next, the point where the mask is created in step 590 is extracted as a region of interest (Crop Region of Interest).

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, and the like alone or in combination. The program instructions recorded on the medium may be specially designed and configured for the embodiment, or may be known and usable to those skilled in computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks and magnetic tapes, optical media such as CDROMs and DVDs, and magnetic-optical media such as floptical disks. A hardware device specially configured to store and execute program instructions such as magneto-optical media and ROM, RAM, flash memory, and the like. Examples of program instructions include not only machine language codes such as those produced by a compiler, but also high-level language codes that can be executed by a computer using an interpreter or the like. The hardware device described above may be configured to operate as one or more software modules to perform the operation of the embodiment, and vice versa.

Hereinafter, an implementation of a system for classifying fruit trees, diseases and pests using a generative adversarial neural network according to an embodiment of the present invention will be described with reference to an example.

* System implementation

In order to implement the system according to an embodiment of the present invention, a Windows 10 Pro operating system may be installed and used in a computer equipped with i5-6500 as a processor, 16G as a memory, and a GeForce GTX 1060 as a graphics card. In addition, CUDA 10.0 is installed to use the GPU, and the system can be implemented using Python3.5 in Juypter Notebook.

1. Implementation of region of interest extraction

7 is a diagram illustrating a result of implementing an ROI.

Referring to FIG. 7, it shows that the original, the region of interest was extracted once, and the region of interest was extracted two times in order from the left. You can see that the region of interest has been properly extracted only when the first row and the image and the second row are performed twice. This occurs while calculating the average value. In the first one, the area of interest is extracted by calculating the average value of the three backgrounds, shadows, and leaves, but in the case of two times, most of the background disappears and the average value of the two shadows and leaves can be obtained. In addition, in FIG. 7, this problem may occur when the focus is blurred at the border of the leaf, although it occurs due to a shadow. Accordingly, in the system according to an embodiment of the present invention, the ROI can be implemented by using twice the ROI extraction so that the ROI is concentrated on a specific part.

2. Implementing generative adversarial neural networks

A block rot is created on a leaf of a blueberry by learning black rot in apples and grapes with the generative hostile neural network of the system according to an embodiment of the present invention. The number of black rot images in apples and grapes is 1441, and this is first learned by the discriminator. After that, the constructor takes a blueberry leaf as an input value and proceeds to learn to generate a black rot-attached leaf.

8 is an exemplary view showing an image of an actual black rot, and FIG. 9 is an exemplary view showing an image of an original blueberry leaf and a blueberry leaf with Black rot generated based on the original blueberry leaf.

Referring to FIGS. 8 and 9, when the pattern of the disease is large, the image generated as the left side of FIG. 9 is awkward, but when the pattern of the disease is small, the leaves with black rot so that the awkward part as shown in the right side of FIG. You can see this is well created. In addition, by limiting the region of interest in the pretreatment process, it can be seen that the location of the disease is properly settled on the leaf.

10 is a table showing classification accuracy of a system according to an embodiment of the present invention compared with an existing system. In other words, FIG. 10 shows Classification Accuracy when using for learning without preprocessing, increasing data through rotation, and using the technique according to an embodiment of the present invention.

Referring to FIG. 10, in the case of using for learning without pre-processing (System 1), a probability of 77.99% came out due to insufficient data set, and in the case of increasing data through rotation (System 2), some data Due to the small amount of three, the probability was 91.14%. In contrast, the Suggest System according to an embodiment of the present invention increases the amount of data used for training by using a generative adversarial neural network, and provides the same amount of data for each class, showing a high numerical accuracy of 98.17%. .

* conclusion

Currently, crops to which smart farms are applied are limited to horticultural or house crops, and there is a problem that a lot of costs are required to introduce smart farms. In addition, the provided data may show a big difference from the actual data, and since the amount of data is different, there is a problem that it must be processed in order to learn, and there is a problem that resources such as a lot of labor are consumed.

To solve this problem, an embodiment of the present invention proposes a method of generating data similar to actual data by learning provided data through a generative adversarial neural network and receiving used data as input values.

The system according to an embodiment of the present invention proposes a preprocessing method to speed up the speed by reducing the amount of computation to reduce the amount of resources used when data is used for learning. Among them, only the objects actually used for learning are extracted by extracting the region of interest. It can lead to learning. After that, a generative adversarial neural network can be used to increase the amount of data used for training and to resolve the data imbalance for each class. Through this, data can be generated so that the smart farm can be applied to various crops, and the accuracy of learning can be increased to 98%.

If data that can be used not only in the classification system used in the system according to an embodiment of the present invention, but also in logistics and automation systems can be generated, not only the classification system but also a wider variety of smart farms can be configured.

Although the embodiments have been described by the limited embodiments and drawings as described above, various modifications and variations can be made from the above description to those of ordinary skill in the art. For example, the described techniques are performed in a different order from the described method, and/or components such as systems, structures, devices, circuits, etc. described are combined or combined in a form different from the described method, or other components Alternatively, even if substituted or substituted by an equivalent, an appropriate result can be achieved.

Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the following claims.

110: network unit
112: GAN network
114: classification network
120: data processing unit
122: data preprocessor
124: data post-processing unit
210: image resizing unit
220: region of interest extraction unit
230: image contrast processing unit

Claims (11)

In order to reduce the resources used when learning data, a data preprocessing unit that accelerates the operation speed by reducing the amount of data calculation by preprocessing the data of the original image including diseases and pests, and storing or filtering the result data of the preprocessing A data processing unit including a data post processing unit; And
A GAN network that generates a similar image corresponding to the result data of the preprocessing through a generative hostile neural network (GAN), and a classification model using the generated similarity image, and an input image based on the learned classification model Network unit including a classification network to determine and classify the types of pests
Fruit tree disease and pest classification system using a generative hostile neural network comprising a.
The method of claim 1,
The data preprocessor
An image size adjusting unit for downsizing by adjusting the size of the original image;
A region of interest extracting unit for extracting a region of interest from the downsized original image; And
Image contrast processing unit to process the extracted image of the region of interest contrast (contrast image)
Fruit tree disease and pest classification system using a generative hostile neural network comprising a.
The method of claim 2,
The region of interest extracting unit
A system for classification of overgrown diseases and pests using a generative adversarial neural network, characterized in that the region of interest is extracted based on an RGB average value calculated using a sobel mask.
The method of claim 1,
The data post-processing unit
A data storage unit that stores the result data of the pre-processing and sends the generated similarity image to the classification network when the generated similarity image is collected in a predetermined amount or more; And
A data filter unit that filters the result data of the pre-processing and filters data whose quality is lower than a predetermined reference value among the generated similar images
Fruit tree disease and pest classification system using a generative hostile neural network comprising a.
The method of claim 1,
The GAN network
Based on a sample image that does not contain pests and pests, the similar image is generated through the generative adversarial neural network.
The method of claim 1,
The GAN network
The process of generating the similar image is repeated until a predetermined predetermined number is reached,
The classification network is
When the number of similar images reaches a certain number, the classification model is trained using the certain number of similar images, thereby increasing the amount of data used for learning and solving the data imbalance for each class. Fruit tree pest classification system using.
The method of claim 1,
The generative adversarial neural network is
All are designed identically except for the existing DCGAN and the discriminator,
The discriminator is
Considering that the original image has a certain shape of a leaf and the subject is a disease occurring in the leaf, the size of the convolution layer in the existing DCGAN is modified from 5*5 to 3*3, and the amount of calculation It is designed to generate similar images of various classes using Sigmoid instead of Relu as an activation function, which saves memory, and uses a generative adversarial neural network.
In the fruit tree pest classification method using a fruit tree pest classification system using a generative hostile neural network,
In order to reduce the resources used when learning data, preprocessing data of the original image including diseases and pests to reduce the amount of data computation to speed up the computation;
Storing or filtering result data of the pre-processing;
Generating a similar image corresponding to the result data of the preprocessing through a generative hostile neural network (GAN);
Learning a classification model using the generated similarity image; And
Determining and classifying the type of disease and pest of the input image based on the learned classification model
Fruit tree disease and pest classification method using a generative hostile neural network comprising a.
The method of claim 8,
Pre-processing the data of the original image to reduce the amount of data computation to speed up the computation speed
Downsizing by adjusting the size of the original image;
Extracting a region of interest from the downsized original image; And
Contrast processing the extracted image of the ROI
Fruit tree disease and pest classification method using a generative hostile neural network comprising a.
The method of claim 8,
The step of generating the similar image
Generating the similar image through the generative adversarial neural network based on the sample image that does not contain pests
Fruit tree disease and pest classification method using a generative hostile neural network comprising a.
The method of claim 8,
The generative adversarial neural network is
All are designed identically except for the existing DCGAN and the discriminator,
The discriminator is
Considering that the original image has a certain shape of a leaf and the subject is a disease occurring in the leaf, the size of the convolution layer in the existing DCGAN is modified from 5*5 to 3*3, and the amount of calculation It is designed to generate similar images of various classes using Sigmoid instead of Relu as an activation function, which reduces memory and saves memory, and classifies fruit diseases and pests using a generative adversarial neural network.

Images