KR102291164B1 - Method for determining weeds with maximum or minimum parameters deep learning structure - Google Patents

Abstract

The present invention relates to a method for determining weeds to which the maximum or minimum number of parameters deep learning structure is applied. It is possible to provide a weed identification method including a determination step of discriminating weeds from a weed dataset using a number deep learning structure.

Description

Method for determining weeds with maximum or minimum parameters deep learning structure to which a deep learning structure is applied

The present invention relates to a method for determining weeds to which the maximum or minimized number of parameters deep learning structure is applied. By selectively using the structure, it relates to a weed identification method to which a deep learning structure is applied to maximize or minimize the number of parameters that can improve the accuracy of image identification of weeds that are difficult to classify and reduce the amount of computation and computation speed.

The world's population is increasing by 1.09% every year, and the demand for resources such as more fabrics, food, water and fuel is also increasing.

It is estimated that by 2050, when the world's population is expected to reach 9 billion people, agricultural production must numerically double than it is today. However, agricultural production growth is hampered by factors that need to be managed, such as climate change, reduced arable land, water scarcity, and infectious diseases and weeds.

Despite decades of efforts to control weeds, farmers are still suffering chronic damage from weeds.

Weeds grow anywhere and for no reason on arable land, and compete with crops for moisture, nutrients, sunlight, etc., and ultimately adversely affect production.

Numerous research results prove that there is a high correlation between a decrease in agricultural production and an increase in weeds. Meanwhile, as interest in healthy food increases, eco-certified or organic crops are in the spotlight.

Accordingly, effective control of weeds is very important and can cause high productivity, eco-friendliness and cost savings for various crops.

Accurate weed identification is important for weed control, because the herbicide used and management method are different for each weed species.

To this end, the research for discrimination using weed images is largely divided into an image processing-based method and a deep learning-based method.

The environment in which weeds actually grow has noise and background that makes it difficult to distinguish the target object.

For this reason, when using a deep learning-based method to accurately classify the species of weeds using a weed image, a large number of weights (w) of the parameters of the deep learning-based prediction model are required. there is a limit

Therefore, a new deep learning that can use a convolutional neural network (CNN) for the accuracy of weed classification and transform the number of weights (w) of parameters involved in prediction in the convolutional neural network model according to the purpose of use We want to develop a structure.

In order to solve the above problems, the present invention is a weed image that is difficult to classify by selectively using a deep learning structure with a maximum number of parameters or a minimum number of parameters so that the number of parameters involved in prediction can be modified according to the purpose of use. An object of the present invention is to provide a weed identification method to which a deep learning structure with a maximizing or minimizing number of parameters that can improve the identification accuracy and reduce the amount of computation and computation speed is applied.

In order to solve the above problems, the weed identification method to which the maximum or minimized number of parameters deep learning structure is applied according to an embodiment of the present invention is a method for weed identification to which the maximum or minimized number of parameters deep learning structure is applied, the maximum number of parameters deep It is possible to provide a weed identification method comprising a determination step of discriminating weeds from a weed dataset using a learning structure or a deep learning structure with a minimum number of parameters.

In addition, before the determination step, it may further include an image processing step of processing the image data into one or more weed datasets.

Here, the image processing step comprises the steps of generating one or more weed images by dividing and cutting leaves, flowers, fruits and outposts that can be seen as weeds (objects) in the image data; Compressing the generated weed image and may include processing the compressed weed image as a weed dataset using a data augmentation technique.

_{In addition, the determination step includes a weight setting step of setting the weights (α n} ) of two or more single classification models from which a probability vector is obtained by inputting a weed dataset when using the deep learning structure with the maximum number of parameters, respectively, and the set weights (α _n) ) may include a mixing step of obtaining a final determination result by mixing.

In addition, the weight setting step is characterized in that the weight (α _n ) is set to 1.

In addition, the weight setting step is characterized in that the weight (α _n ) is calculated and set through Equation 1 below.

[Equation 1]

(Where n is the number of single classification models, and MAX is the maximum value among the probability values for each final class of the single classification model.)

In addition, the weight setting step is characterized in that the weight (α _n ) is calculated and set through a multi-layer perceptron.

In addition, the weight setting step is characterized in that the probability vector of each single classification model is converted into a confidence value (real value) that is an input value of the multi-layer perceptron through an entropy method or a maximum difference method.

In addition, the mixing step is characterized in that the final determination result is obtained by mixing the _{weights (α n} ) of each single classification model through a weighted linear combination method or a weighted squared combination method.

In addition, in the determining step, when using the deep learning structure of the minimum number of parameters, the classifier through the family determination step of determining a family by inputting a weed dataset to the parent classifier and a child classifier corresponding to the determined family ) may include a species identification step of obtaining a final determination result by determining the

In addition, the species determination step is characterized in that the final determination result is obtained through the following Equation 6 using the values obtained from the parent classifier and the child classifier.

[Equation 6]

는 자식분류기의 예측값,

는 부모분류기의 예측값, P는 부모분류기의 예측값에서의 자식분류기의 예측값의 확률이다)(where W is the weight of a single classification model,

is the predicted value of the subclassifier,

is the predicted value of the parent classifier, and P is the probability of the predicted value of the child classifier in the predicted value of the parent classifier)

In addition, the weed dataset is characterized in that it has information about the weed family (family) and species (class) in a tree structure.

In addition, the determining step may further include a predictive value determining step of determining whether the parent classifier correctly predicted the correct answer or incorrect using a threshold value in the over-determining step.

에서 0으로 수렴하고, 부모분류기가 정답과 틀리게 예측할 경우 해당되는 하나 이상의 자식분류기의 정답 확률 벡터를 계산하여 가장 높은 값을 가지는 예측값을 도출하는 것을 특징으로 한다.In addition, the determining step further includes a reinforcement step of finding a correct answer when the parent classifier incorrectly predicts the correct answer in the predicted value determination step, and the reinforcement step is performed according to the determination in the predicted value determination step through the following Equations 7 and 8, If the parent classifier correctly predicts the correct answer

converges to 0, and when the parent classifier predicts wrongly from the correct answer, it is characterized in that the prediction value having the highest value is derived by calculating the correct probability vector of the corresponding one or more child classifiers.

[Equation 7]

[Equation 8]

(where Y is the child classifier, n is the number of the child classifier, I is the weed dataset, and F _i is the parent classifier)

In addition, in the determination step, when using the deep learning structure of the minimum number of parameters, the final determination result is obtained through the following Equation 9 using the parent classifier and the child classifier, and the family and the class are simultaneously identified. characterized by discriminating

[Equation 9]

는 j번째 과의 k번째 종에 대한 확률값,

는 자식분류기 Y_k의 예측값,

는 부모분류기 F_j의 예측값이다)(here,

is the probability value for the kth species of the jth family,

is the predicted value of the subclassifier _{Y k ,}

is the predicted value of the parent classifier F _{j )}

The weed identification method to which the maximization or minimization number of parameters deep learning structure is applied according to an embodiment of the present invention as described above is a deep learning structure or minimizes the number of parameters so that the number of parameters involved in prediction can be modified to suit the purpose of use. By selectively using the learning structure, it is possible to improve the accuracy of weed image identification, which is difficult to classify, and reduce the amount of computation and computation speed.

In other words, by using two or more single classification models together through the deep learning structure to maximize the number of parameters, the accuracy of weed identification can be improved compared to when a single classification model is used, and as the number of single classification models used increases, It can be made to show high accuracy.

In addition, by adjusting the number of parameters through the minimized number of parameters deep learning structure, the amount of calculation and the calculation speed can be drastically reduced while the accuracy can be set to a certain level or higher.

Accordingly, in places where high computing power such as supercomputers or servers are supported, the maximum number of parameters can be used to further improve accuracy by using deep learning structures, and in places where computing power is weak such as embedded systems, the minimum number of parameters can be deep A learning structure can be used to effectively identify weeds.

1 is a flowchart illustrating a method for determining weeds to which a deep learning structure is applied to maximize or minimize the number of parameters according to an embodiment of the present invention.
FIG. 2 is a flowchart schematically illustrating step S10 of FIG. 1 .
3 is a conceptual diagram illustrating a deep learning structure with a maximizing number of parameters used in a method for determining weeds to which a maximizing or minimizing number of parameters deep learning structure is applied according to an embodiment of the present invention.
4 is a flowchart schematically illustrating step S21 of FIG. 1;
Figure 5 is a conceptual diagram showing the structure of a multi-layer perceptron (MLP) used in the maximum number of parameters deep learning structure of the weed discrimination method to which the maximum or minimized number of parameters deep learning structure is applied according to an embodiment of the present invention.
6 is a conceptual diagram illustrating a deep learning structure with a minimized number of parameters used in a method for determining weeds to which a deep learning structure is applied to maximize or minimize the number of parameters according to an embodiment of the present invention.
7 is a flowchart schematically illustrating step S22 of FIG. 1;
8 is a histogram showing the accuracy according to the prediction value of the parent classifier to set the threshold value used in the reinforcement deep learning structure.
Figure 9 is a conceptual diagram illustrating a reinforcement deep learning structure used in the minimum number of parameters deep learning structure of the weed identification method to which the maximum or minimized number of parameters deep learning structure is applied according to an embodiment of the present invention.
Figure 10 is a conceptual diagram illustrating a deep learning structure of the minimum number of parameters of end-to-end learning used in the method for determining weeds to which the maximum or minimized number of parameters deep learning structure is applied according to an embodiment of the present invention.

Hereinafter, the description of the present invention with reference to the drawings is not limited to specific embodiments, and various modifications may be made and various embodiments may be provided. In addition, it should be understood that the contents described below include all conversions, equivalents, and substitutes included in the spirit and scope of the present invention.

In the following description, terms such as first and second are terms used to describe various components, meanings are not limited thereto, and are used only for the purpose of distinguishing one component from other components.

Like reference numbers used throughout this specification refer to like elements.

As used herein, the singular expression includes the plural expression unless the context clearly dictates otherwise. In addition, terms such as "comprises", "comprising" or "have" described below are intended to designate the existence of features, numbers, steps, operations, components, parts, or combinations thereof described in the specification. It should be construed as not precluding the possibility of the presence or addition of one or more other features, numbers, steps, operations, components, parts, or combinations thereof.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. Terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with the meaning in the context of the related art, and should not be interpreted in an ideal or excessively formal meaning unless explicitly defined in the present application. does not

In addition, in the description with reference to the accompanying drawings, the same components are given the same reference numerals regardless of the reference numerals, and the overlapping description thereof will be omitted. In describing the present invention, if it is determined that a detailed description of a related known technology may unnecessarily obscure the gist of the present invention, the detailed description thereof will be omitted.

Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings 1 to 10.

1 is a flowchart illustrating a method for determining weeds to which a deep learning structure is applied to a maximizing or minimizing number of parameters according to an embodiment of the present invention, FIG. 2 is a flowchart schematically illustrating step S10 of FIG. 1, and FIG. 3 is the present invention A conceptual diagram illustrating a deep learning structure with a maximized number of parameters used in a weed identification method to which the maximum or minimized number of parameters deep learning structure is applied according to an embodiment, FIG. 4 is a flowchart schematically illustrating step S21 of FIG. 5 is a conceptual diagram illustrating the structure of a multi-layer perceptron (MLP) used in the maximizing parameter number deep learning structure of the weed identification method to which the maximizing or minimizing parameter number deep learning structure is applied according to an embodiment of the present invention.

The present invention relates to a method for determining weeds to which the maximum or minimum number of parameters deep learning structure is applied, so that the number of parameters involved in prediction can be modified according to the purpose of use. By using it as a , it is intended to improve the accuracy of weed image identification, which is difficult to classify, and reduce the amount of computation and computation speed.

Here, the maximizing number of parameters deep learning structure can be used where high computing power is provided, so that accuracy can be further improved. It can be reduced to have more than a certain level of accuracy. However, the present invention is not limited thereto and may be selectively used according to the purpose of use.

Referring to FIG. 1 , a method for determining a weed to which a deep learning structure is applied with a maximizing or minimizing number of parameters according to an embodiment of the present invention (hereinafter referred to as a 'weed determination method') may include an image processing step (S10) and a determination step have.

Here, the determination step is a determination step (S21) of discriminating weeds from the weed dataset using the deep learning structure with the maximum number of parameters, or the weeds from the weed dataset using the minimum number of parameters deep learning structure, depending on the deep learning structure used. It may include a determination step (S22) of determining.

First, a detailed description will be given of a method for determining weeds through step S21.

Specifically, the image processing step (S10) is a step of processing image data in which one or more weeds are photographed into one or more weed datasets, and weeds can be determined using the image data as it is, but the resolution is lowered and the object (weeds) This is a step to improve discrimination accuracy by solving other noises, out-of-focus abnormalities, etc. to enable smooth learning.

Referring to Figure 2, step S10 may include generating a weed image (S100), compressing the weed image (S101) and processing the weed dataset (S102).

The step of generating a weed image (S100) may generate one or more weed images by classifying leaves, flowers, fruits and outposts that can be seen as weeds (objects) in high-resolution image data and cutting them according to characteristics.

Compressing the weed image (S101) can be made into a desirable form for input to a single classification model by compressing each of the one or more weed images generated in step S100.

Here, the weed image can be compressed according to the defined standard horizontally, vertically, and in depth, preferably compressed in a square shape, but is not limited thereto.

In addition, in step S101, if the weed image is compressed into a square, the characteristic of the length or ratio for the weed species may disappear due to the characteristics of the weed. It can be filled via values (eg zero values).

Processing as a weed dataset (S102) may generate one or more weed datasets by processing one or more weed images compressed in step S101 as a weed dataset using a data augmentation technique, respectively.

This is because even if weed images are compressed in step S101, all weed images that exist in the world cannot be acquired. it can be

More specifically, step S102 can be generated as a final weed dataset by processing the weed image by using one or more methods of left and right inversion, vertical inversion, three-dimensional twist, movement and rotation as a data augmentation technique.

In this case, the movement and rotation method may be performed by treating empty pixels as 0.

The determination step (S21) can determine the weeds from one or more weed datasets using the deep learning structure of the maximum number of parameters.

The maximum number of parameters The deep learning structure is implemented as shown in FIG. 3, and it is intended to improve the accuracy by using the maximum number of parameters. Such a deep learning structure, for example, can be said to be a structure that gathers several experts (single classification model) to identify weeds, sees images of weeds, gives individual opinions, and synthesizes them.

That is, by using two or more single classification models together, the accuracy of weed identification can be improved than when using one single classification model, and as the number of single classification models used increases, the accuracy becomes higher. can do.

Here, the single classification model is a convolutional neural network model (CNN), which may be a VGGNet model, a ResNet model, an Inception-Resnet model, a MobileNet model, a NASNet model, or a MobileNetV2 model, but is not limited thereto.

Referring to FIG. 4 , step S21 may include a learning step S210 , an input step S211 , a weight setting step S212 , and a mixing step S213 .

The learning step (S210) is a step of learning two or more single classification models, and by learning weed data, it is possible to discriminate weeds when a weed dataset is input.

Here, the two or more single classification models may be different models, but is not limited thereto.

In addition, the weed data may be a plurality of images classified into families and species, but is not limited thereto, and may be a plurality of images classified only by species.

At this time, the image used for the weed data may be generated by processing the same as the step S10 of generating the weed data set.

For example, weed data may be configured as shown in Table 1 below, which is not limited thereto because it is merely an example.

In addition, in step S210, before learning the weed data, transfer learning may be performed with 1000 image net taxa.

This is to make learning on weed data more efficient by pre-learning with 1000 taxa on ImageNet because 339000 images out of 1000 taxa are related to plants and flowers.

In the input step (S211), a probability vector can be obtained by inputting and learning each weed dataset to two or more single classification models from which weed data is learned in step S210.

_{The weight setting step (S212) may set the weights (α n} ) of two or more single classification models from which the probability vectors are obtained by inputting the weed dataset in the step S211, respectively. Here, calculating the weights for each single classification model is to increase the accuracy by calculating the weights of each single classification model, since an error may increase if the weights of a single classification model having different probability vectors are all the same.

In step S212 , the weight α _n may be set to 1 or a number other than 1.

In step S212, when the weight α _{n is set to a number other than 1, the weight α n} may be set through a non-learning or learning method.

Specifically, when the weight α _n is set through the non-learning method, step S212 may be divided into two methods.

First, in step S212, an arbitrary value may be determined at random, and the arbitrary value may be set _{as the weight (α n ).}

Second, in step S212, the weight α _n may be calculated and set through Equation 1 below.

[Equation 1]

Here, n is the number of single classification models, and MAX is the maximum value among the probability values for each final class of the single classification model.

In addition, step S212 may be set by calculating a weight (α _n) with a weighting (α _n), is also a multi-layered perceptron (Multi-Layer Perceptron) having a structure such as a 5 when set through learning. This makes it possible to use an optimized weight through learning rather than allowing the user to select the weight α _{n .}

In order to apply such a multi-layer perceptron, a confidence value (real value), which is a single vector, is required, and the confidence value (real value) can be obtained through the entropy method or the maximum difference method.

The confidence value is a value for measuring the reliability of a probability vector before making a final classification decision. For example, if there is a probability vector with A = [0.900;0.050;0.020; 0.015; 0.015] and B = [0.500;0.400;0.050;0.030; Although guaranteed, even if the maximum value of 0.500 is selected in B, there is no significant difference from the next value of 0.400, so the reliability may be lowered.

In the case of B, it is desirable to increase the reliability by changing the weights. It is difficult to modify in the method in which the user selects the weights (non-learning method), but when setting the weights by learning through the multi-layer perceptron, There is an advantage that optimization by changing the weight is possible.

That is, in step S212, when using a multi-layer perceptron, the probability vector of each single classification model is converted into a confidence value (real value) that is an input value of the multi-layer perceptron through the entropy method or the maximum difference method. , a multi-layer perceptron can be applied.

Here, when using the entropy method, a confidence value (real value) can be obtained through Equation 2 below.

[Equation 2]

Here, p _i is a probability vector and C(p _i ) is a confidence value.

In addition, when the maximum difference method is used, a confidence value (real value) can be obtained through Equation 3 below.

[Equation 3]

Here, p _i is a probability vector and C(p _i ) is a confidence value.

The mixing step (S213) may be when the weight (α _n) set by the step S212, a mixture of a weight (α _n) of each single classification model obtain a final discrimination result.

In step S213, a final determination result may be obtained by mixing the _{weights α n} of each single classification model through a weighted linear combining method or a weighted squared combining method.

Here, when the weighted linear combining method is used, the weights α _n can be mixed through Equation 4 below.

[Equation 4]

where B is a matrix, P is a set of probability vectors, W is a set of weight priorities, S is a score vector summing all elements in a column of B, S _i is a score corresponding to class i, n is a single classification model is the number

In this case, B is a matrix having a size of c × n, and c is a length, which may be determined by the classification result.

In addition, when the weighted square combining method is used, the weights α _n can be mixed through Equation 5 below.

[Equation 5]

where B is a matrix, P is a set of probability vectors, W is a set of weight priorities, S is a score vector summing all elements in a column of B, S _i is a score corresponding to class i, n is a single classification model is the number

In this case, B is a matrix of c × n size, c is a length, and may be a product of all elements in a column of B.

Accuracy can be improved by determining the weeds of the weed dataset through the above method.

Hereinafter, a method for determining weeds through step S22 will be described in detail.

6 is a conceptual diagram illustrating a deep learning structure with a minimized number of parameters used in a method for determining a weed to which a maximizing or minimizing number of parameters deep learning structure is applied according to an embodiment of the present invention, and FIG. 7 schematically shows step S22 of FIG. 8 is a histogram showing the accuracy according to the prediction value of the parent classifier to set the threshold value used in the reinforcement deep learning structure, and FIG. 9 is the number of maximization or minimization parameters deep learning structure according to an embodiment of the present invention is a conceptual diagram showing a reinforcement deep learning structure used in a deep learning structure, the number of minimized parameters of the weed identification method applied It is a conceptual diagram showing the deep learning structure with the minimum number of parameters for end-to-end learning.

As shown in FIG. 1 , after step S10, step S22 may be performed.

Determination step (S22) can determine the number of weeds from one or more weed datasets using a deep learning structure to minimize the number of parameters.

The number of minimized parameters The deep learning structure is implemented as shown in FIG. 6, and it is intended to reduce the amount of computation and the computation speed by using the number of minimized parameters, but to secure a certain level of accuracy or more.

Weeds can be classified taxonomically into families and classes. Species with similar characteristics are grouped into families. There are many features.

The minimum number of parameters deep learning structure uses this, for example, an expert (single classification model - single classification model - It can be said that the subclassifier) is a structure for discriminating species.

Through this, the number of parameters can be minimized by allowing the discrimination to be made only with the weed data for the species of the relevant family rather than using the weed data for all species. can make it more than that.

Referring to FIG. 7 , step S22 may include a learning step ( S220 ), a fruit determination step ( S221 ), and a species determination step ( S224 ).

The learning step (S220) is a step of learning a single classification model, classifying weed data into family and species data, and then learning the family data among weed data in a single classification model as a parent classifier and among weed data in a single classification model. Species data can be trained individually for each species to make one or more subclassifiers.

The single classification models used herein may be the same model, but are not limited thereto.

In addition, in step S220, before learning over-data or medium-data, transfer learning may be performed on a single classification model with 1000 image net taxa.

The family determination step (S221) may determine a family (family) by inputting a weed dataset to the parent classifier.

In the species determination step S224, a final determination result may be obtained by determining a species through a child classifier corresponding to the determined family.

In this case, in step S224, a final determination result may be obtained through Equation 6 below using the values obtained from the parent classifier and the child classifier.

[Equation 6]

는 자식분류기의 예측값,

는 부모분류기의 예측값, P는 부모분류기의 예측값에서의 자식분류기의 예측값의 확률이다.where W is the weight of a single classification model,

is the predicted value of the subclassifier,

is the predicted value of the parent classifier, and P is the probability of the predicted value of the child classifier in the predicted value of the parent classifier.

In order to discriminate weeds through the deep learning structure of such a minimized number of parameters, the weed dataset must have information on the family and class of weeds in a tree structure.

In addition, the determination step (S22) may further include a prediction value determination step (S222) after the and determination step (S221).

The prediction value determination step S222 may determine whether the parent classifier correctly predicted the correct answer in the over determination step S221 using a threshold value.

The threshold value used here may be set through a histogram showing the accuracy according to the prediction value of the parent classifier as shown in FIG. 8, and may be set to 99%, but is not limited thereto.

In addition, the determination step S22 may further include a reinforcement step S223 when the parent classifier incorrectly predicts the correct answer in step S222 after the prediction value determination step S222.

The reinforcement step S223 applies a structure that rewards when the misrecognized parent classifier is predicted as shown in FIG. 9, and when the parent classifier incorrectly predicts the correct answer in step S222, the correct answer can be found.

In step S223, according to the determination of step S222, if the parent classifier predicts wrongly from the correct answer through the following equations 7 and 8, the prediction value having the highest value can be derived by calculating the correct answer probability vector of one or more child classifiers. .

[Equation 7]

[Equation 8]

Here, Y is the child classifier, n is the number of the child classifier, I is the weed dataset, and F _i is the parent classifier.

에서 0으로 수렴시켜, 원래대로 판단된 과에 해당되는 자식분류기만을 사용하게 할 수 있다.In step S223, if it is determined that the parent classifier correctly predicted the correct answer in step S222,

By converging to 0, it is possible to use only the child classifiers corresponding to the originally judged family.

Referring to FIG. 10 , in the determination step (S22) of discriminating weeds from one or more weed datasets using a deep learning structure with a minimized number of parameters, as shown in FIG. 7 , the parent classifier and the child classifier are hierarchically divided In addition to being processed, it is also possible to determine a family and a class at the same time as it consists of a deep learning structure with a minimum number of parameters for end-to-end learning.

This is to obtain only the probability value from the parent classifier and apply it directly to the child classifier, so that the family and the class can be discriminated at the same time.

Specifically, in step S22 to which end-to-end learning is applied, a final determination result may be obtained through Equation 9 below using a parent classifier and a child classifier.

[Equation 9]

는 j번째 과의 k번째 종에 대한 확률값,

는 자식분류기 Y_k의 예측값,

는 부모분류기 F_j의 예측값이다.here,

is the probability value for the kth species of the jth family,

is the predicted value of the subclassifier _{Y k ,}

is the predicted value of the parent classifier F _{j .}

By determining the weeds of the weed dataset through the above method, it is possible to reduce the amount of computation and the computation speed.

Hereinafter, the present invention will be described in more detail by way of Examples.

However, the following examples are merely illustrative of the present invention, and the content of the present invention is not limited by the following examples.

[ Example ] Number of maximization or minimization parameters deep learning Check structure usability

In order to verify the effect of the present invention, the number of maximization and minimization parameters of a convolutional neural network single classification model and the present invention by processing image data for 5 weeds and 21 species as a weed dataset and then inputting it to determine weeds Methods using deep learning structures were compared and analyzed.

At this time, the single classification model trained weed data in the state of transfer learning of 1000 taxa on ImageNet, and the weed dataset is a color image and the size is 224*224.

Weed data used for learning are shown in Table 1.

1) single classification model

As a result, the accuracy and number of parameters for each of the convolutional neural network single classification models (VGGNet, ResNet, Inception-Resnet, MobileNet, NASNet, MobileNetV2) 21 species were shown in Table 2 below.

2) Maximize parameter deep learning structure

In the case of the maximization parameter deep learning structure, NASNet, ResNet, VGGNet, Inception-Resnet, and MobileNet were used in combination in the order of high accuracy, referring to the single classification model result. At this time, MobileNetV2 was not used because it was a model overlapping with MobileNet.

The accuracy of the maximization parameter deep learning structure according to the number of single classification models was analyzed, and the results of the combination showing the best accuracy in the number of single classification models among all combinations that can be used in the maximization parameter deep learning structure are shown in Table 3 below.

As shown in Table 3 above, it was confirmed that the accuracy of the maximization parameter deep learning structure using two or more single classification models was higher than that of the NASNet model with the highest accuracy in a single classification model, and as the number of single classification models increases, the accuracy could be seen to improve.

3) Minimization parameters deep learning structure

<Minimized parameter deep learning structure>

For the minimization parameter deep learning structure, MobileNetV2 with a small number of parameters was used by referring to the single classification model result. At this time, weed data learning was used by classifying weed data into family data and class data, a parent classifier that learned the family data, and five child classifiers that learned species data for each species.

Through this, the subclassifier that discriminates species has a structure that classifies up to 11 classes, and the structure is reduced by half compared to the 21 classes that had to be classified in the existing single classification model.

The mobileNetV2-based minimization parameter deep learning structure implemented in this way is shown in Table 4, and it was compared with the ResNet model.

As can be seen from Table 4, the ResNet model showed 98.0% accuracy while using the number of parameters 21,298,010, and the minimized parameter deep learning structure used the number of parameters 4,897,010. It was confirmed that the accuracy was 97.23%.

Although the accuracy was less than 0.7%, there was a slight difference, but it was confirmed that the number of parameters was reduced by more than 3 times.

<Minimum parameter deep learning structure with reinforcement deep learning structure applied>

It is determined whether the parent classifier predicted the correct answer by using the threshold value in the minimization parameter deep learning structure. A structure to find the correct answer was applied by deriving a predicted value with .

As a result, the number of parameters and accuracy were shown in Table 5 below.

As can be seen from Table 5 above, it was confirmed that the minimization parameter deep learning structure to which the reinforcement deep learning structure is applied also reduces the number of parameters by more than 3 times compared to the ResNet model, while improving the accuracy somewhat compared to the existing minimization parameter deep learning structure.

<Deep learning structure with minimal parameters of end-to-end learning>

The number of parameters and the accuracy were compared and analyzed by applying the parameter deep learning structure to minimize the end-to-end learning. The results are shown in Table 6 below.

As can be seen from Table 6, it was confirmed that the accuracy was slightly reduced compared to the existing minimized parameter deep learning structure, but the number of parameters was further reduced by about 3 times, and it was confirmed that it was reduced by about 10 times than the ResNet model. .

Accordingly, it is determined that when using the minimized parameter deep learning structure, the number of parameters can be minimized to reduce the amount of computation and the computation speed.

The embodiment of the present invention described above is not implemented only through an apparatus and/or method, but may be implemented through a program for realizing a function corresponding to the configuration of the embodiment of the present invention, a recording medium in which the program is recorded, etc. Also, such an implementation can be easily implemented by those skilled in the art to which the present invention pertains from the description of the above-described embodiments.

Although the embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improved forms of the present invention are also provided by those skilled in the art using the basic concept of the present invention as defined in the following claims. is within the scope of the right.

Claims (15)

In the weed identification method to which the deep learning structure is applied, the number of maximizing or minimizing parameters,
Including a discrimination step of discriminating weeds from a weed dataset using the maximizing number of parameters deep learning structure,
The determination step is
A weight setting step of setting the weights _{(α n} ) of two or more single classification models from which probability vectors are obtained by inputting a weed dataset, respectively;
Including a mixing step of mixing the set weight (α _n ) to obtain a final determination result,
The weight setting step is
It is set by calculating the _{weight (α n} ) through the multi-layer perceptron,
Weed discrimination method, characterized in that by converting the probability vector of each single classification model into a confidence value (real value) that is an input value of the multi-layer perceptron through the entropy method or the maximum difference method.
According to claim 1,
Before the determination step,
Weed identification method further comprising an image processing step of processing the image data into one or more weed datasets.
3. The method of claim 2,
The image processing step is
generating one or more weed images by dividing and cutting leaves, flowers, fruits, and outposts that can be seen as weeds (objects) from the image data;
Compressing the generated weed image and
Weed identification method comprising the step of processing the compressed weed image as a weed dataset using a data augmentation technique.
delete delete delete delete delete According to claim 1,
The mixing step is
Weed discrimination method, characterized in that by mixing the _{weights (α n} ) of each single classification model through a weighted linear combining method or a weighted squared combining method to obtain a final discrimination result.
In the weed identification method to which the deep learning structure is applied, the number of maximizing or minimizing parameters,
Including a discrimination step of discriminating weeds from a weed dataset using a deep learning structure with a minimum number of parameters,
The determination step is
A family determination step of determining a family by inputting a weed dataset to the parent classifier;
a species identification step of obtaining a final determination result by determining a species through a child classifier corresponding to the determined family;
A prediction value determination step of judging whether the parent classifier correctly predicted the correct answer or wrong using a threshold in the over-determining step, and
a reinforcing step of finding the correct answer when the parent classifier incorrectly predicts the correct answer in the predictive value determination step;
The species identification step is
A final determination result is obtained through Equation 6 below using the values obtained from the parent classifier and the child classifier,
The reinforcement step is
When the parent classifier accurately predicts the correct answer through the following Equations 7 and 8 according to the determination in the prediction value determination step

converges to 0, and when the parent classifier predicts wrongly from the correct answer, the method for determining the weed, characterized in that the prediction value with the highest value is derived by calculating the correct probability vector of the corresponding one or more child classifiers.
[Equation 6]

(where W is the weight of a single classification model,

is the predicted value of the subclassifier,

is the predicted value of the parent classifier, P is the probability of the predicted value of the child classifier in the predicted value of the parent classifier)
[Equation 7]

[Equation 8]

(where Y is the child classifier, n is the number of the child classifier, I is the weed dataset, and F _i is the parent classifier)
delete 11. The method of claim 10,
The weed dataset is,
Weed identification method, characterized in that it has information about the weed family (family) and species (class) in a tree structure.
delete delete In the weed identification method to which the deep learning structure is applied, the number of maximizing or minimizing parameters,
Including a discrimination step of discriminating weeds from a weed dataset using a deep learning structure with a minimum number of parameters,
The determination step is
By using the parent classifier and the child classifier to obtain the final determination result through Equation 9 below,
Weed identification method, characterized in that the identification of the family (family) and species (class) at the same time.
[Equation 9]

(here,

is the probability value for the kth species of the jth family,

is the predicted value of the subclassifier _{Y k ,}

is the predicted value of the parent classifier F _{j )}

Images