KR20200121206A - Teacher-student framework for light weighted ensemble classifier combined with deep network and random forest and the classification method based on thereof - Google Patents

Abstract

The present invention relates to a teacher-student framework for lightening an ensemble classifier in which a deep network and random forest are combined, and a classification method based on the same. The teacher-student framework comprises: a teacher training module (100) for training a teacher model formed of a teacher deep network and a teacher random forest by using data set A; a soft target data generation module (200) for inputting data set B into the teacher deep network and the teacher random forest trained in the teacher training module (100), and combining two outputs to generate soft target data set B′; a student training module (300) for training a student model formed of a student network and a student random forest by using the data set B′ generated in the soft target data generation module (200); and a classification module (400) for performing classification by combining two outputs of the student network and the student random forest trained in the student learning module (300).

Description

TEACHER-STUDENT FRAMEWORK FOR LIGHT WEIGHTED ENSEMBLE CLASSIFIER COMBINED WITH DEEP NETWORK AND RANDOM FOREST AND THE CLASSIFICATION METHOD BASED ON THEREOF }

The present invention relates to a teacher-student framework and a classification method based on the same, and more specifically, to a teacher-student framework for lightening an ensemble classifier combined with a deep network and a random forest, and a classification method based on the same. About.

Neural Network is a structure adopted by computers to solve problems in a way similar to the way humans handle problems through the brain.When neurons as a mathematical model are interconnected to form a network, it is a neural network or artificial It is called an Artificial Neural Network.

Since the neural network acts as a processor for each neuron to operate independently, it has excellent parallelism, and because information is distributed across many connection lines, even if a problem occurs in some neurons, it does not affect the entire system, so fault tolerance. tolerance), and the ability to learn about a given environment. Due to such characteristics, it is used for problem solving in the field of artificial intelligence, and is used in various fields such as character recognition, speech recognition, classification, diagnosis, and prediction.

Among various deep neural networks, Convolutional Neural Network (CNN) has high accuracy and is widely used in various fields such as pedestrian detection, pose orientation estimation (POE) based on computer vision, and voice recognition. have. CNNs require a large number of data sets for training and testing, depending on their application. In addition, since the amount of calculation is large, there is a limitation in that a large-scale high-level computing device is required compared to a conventional classifier. Therefore, in fields that require rapid and accurate classification, such as POE, there is a need to develop a method capable of reducing processing time and memory amount while ensuring high performance.

On the other hand, Random Forest in Machine Learning is a kind of ensemble learning method used for classification and regression analysis, and it is a class (classification) or average predicted value (regression analysis) from a number of decision trees constructed in the training process. It works by outputting ). Random forest is an ensemble method of randomly learning several decision trees. The random forest method largely consists of a learning step that constructs a number of decision trees and a test step that classifies or predicts when an input vector is received. Random forests are used in various applications such as detection, classification, and regression.

The parameters that have the greatest influence in a random forest are the size of the forest (number of trees) and the maximum allowable depth. Among them, the size of the forest (the number of trees) is a parameter that determines how many trees the total forest consists of. If the size of the forest is small, that is, if the number of trees is small, the time taken to construct and test the trees is short, but the generalization ability is low, and there is a high probability of producing incorrect results for any input data point. On the other hand, if the size of the forest is large, that is, if the number of trees is large, high performance is guaranteed, but there are disadvantages of lengthening training and testing time and increasing the amount of memory. Therefore, there is a need to develop an improved random forest method capable of reducing processing time and memory amount while ensuring high performance.

On the other hand, as a prior art related to the present invention, Patent No. 10-1901307 (Name of the invention: Class classification method, apparatus and computer-readable recording medium using a deep neural network based on weighted fuzzy membership function), Patent Publication No. 10-2018 No. -0046122 (title of the invention: a method and apparatus for generating a toxicity prediction model of oxide nanomaterials using a random forest technique) and the like have been disclosed.

The present invention has been proposed to solve the above problems of the previously proposed methods, and develops a new ensemble classifier by combining a deep network and a random forest, and a soft target data set B ^* which is an output of the teacher model as an input. By learning the student model, the ensemble classifier developed through the teacher-student framework consisting of the teacher model and the student model can be lightened and output more flexible classification results. Data set A and class including class labels A teacher-student framework for lightening the ensemble classifier combined with deep network and random forest, which can prevent overfitting of the teacher model by training the teacher model using the label-free dataset B, and Its purpose is to provide a classification method based on this.

A teacher-student framework for lightening the weight of an ensemble classifier combined with a deep network and a random forest according to the features of the present invention for achieving the above object,

As a teacher-student framework,

A teacher learning module for learning a teacher model composed of a teacher deep network and a teacher random forest using the data set A;

A soft target data generation module for inputting data set B into the deep teacher network and teacher random forest learned in the teacher learning module, and combining the output two outputs to generate a soft target data set B ^* ;

A student learning module for learning a student model composed of a student network and a student random forest by using the data set B ^* generated by the soft target data generation module; And

It is characterized in that it comprises a classification module that performs classification by combining two outputs of the student network and the student random forest learned in the student learning module.

Preferably, the data set A,

It may be a hard target data set including a class label.

Preferably, the data set B,

It may be a data set that does not include a class label.

Preferably, the teacher learning module,

A first teacher learning unit for learning the deep teacher network using the data set A; And

It may include a second teacher learning unit for learning a teacher random forest by using a feature map of the deep teacher network.

Preferably, the deep teacher network,

In order to obtain a soft target output that is a probability value of each class, a softened softmax function can be applied.

Preferably,

A pre-processing module for performing pre-processing on the input image by applying wavelet transform may be further included.

Preferably, the student learning module,

A first student learning unit for learning a student network by using the data set B ^* generated by the soft target data generation module; And

It may include a second student learning unit for learning a student random forest by using the data set B ^* generated by the soft target data generation module.

More preferably, the first student learning unit,

(3-1-1) initializing the parameter (W _S ) of the student network;

(3-1-2) inputting the data set B ^* into a pre-trained network;

(3-1-3) calculating a loss function (L(Ws));

를 W_S ^*로 업데이트하는 단계; 및(3-1-4)

Updating to W _S ^* ; And

(3-1-5) By performing the step of selecting the optimal parameter W _S ^* for the student network, the student network may be trained.

Even more preferably, in the step (3-1-3),

The loss function can be calculated using the following equation.

In the above equation, N is the number of samples in the data set B ^* , C is the number of classes, and P _T (x _i |c _j ) and P _S (x _i |c _j ) are the teachers and students for the input vector x _i , respectively. This is the posterior class probability.

More preferably, the second student learning unit,

(3-2-1) copying the t-th tree structure (T-RF _t ) of the teacher random forest to the t-th tree structure (S-RF _t ) of the student random forest to perform transfer learning;

(3-2-2) inputting a data set B ^* having an input vector v into one of the decision trees of S-RF _t ;

(3-2-3) generating a split function f(v) at node O;

(3-2-4) calculating an information gain ΔE at node O;

(3-2-5) calculating cross-entropy Tr(Te, S) _t ; And

(3-2-6) If the calculated cross-entropy is less than the minimum threshold for stopping boosting, the current S-RF _t is stored, and otherwise, the step of re-performing from step (3-2-2) is T It is possible to learn the student random forest by executing until a random decision tree is constructed.

A classification method based on a teacher-student framework for lightening an ensemble classifier combined with a deep network and a random forest according to a feature of the present invention for achieving the above object,

As a classification method based on the teacher-student framework,

(1) using the data set A, training a teacher model composed of a deep teacher network and a teacher random forest;

(2) inputting the data set B into the deep teacher network and the teacher random forest learned in step (1), and combining the two outputs to generate a soft target data set B ^* ;

(3) learning a student model consisting of a student network and a student random forest by using the data set B ^* generated in step (2); And

(4) It is characterized in that it comprises the step of performing classification by combining the two outputs of the student network and the student random forest learned in step (3).

Preferably, the data set A,

It may be a hard target data set including a class label.

Preferably, the data set B,

It may be a data set that does not include a class label.

Preferably, the step (1),

(1-1) learning the deep teacher network using the data set A; And

(1-2) It may include the step of learning a teacher random forest using a feature map of the deep teacher network.

Preferably, the deep teacher network,

In order to obtain a soft target output that is a probability value of each class, a softened softmax function can be applied.

Preferably,

(0) The step of performing preprocessing on the input image by applying the wavelet transform may further be included.

Preferably, the step (3),

(3-1) learning a student network by using the data set B ^* generated in step (2); And

(3-2) Using the data set B ^* generated in step (2), it may include the step of learning a student random forest.

More preferably, the step (3-1),

(3-1-1) initializing the parameter (W _S ) of the student network;

(3-1-2) inputting the data set B ^* into a pre-trained network;

(3-1-3) calculating a loss function (L(Ws));

를 W_S ^*로 업데이트하는 단계; 및(3-1-4)

Updating to W _S ^* ; And

(3-1-5) By performing the step of selecting the optimal parameter W _S ^* for the student network, the student network may be trained.

Even more preferably, in the step (3-1-3),

The loss function can be calculated using the following equation.

In the above equation, N is the number of samples in the data set B ^* , C is the number of classes, and P _T (x _i |c _j ) and P _S (x _i |c _j ) are the teachers and students for the input vector x _i , respectively. This is the posterior class probability.

More preferably, the step (3-2),

(3-2-1) copying the t-th tree structure (T-RF _t ) of the teacher random forest to the t-th tree structure (S-RF _t ) of the student random forest to perform transfer learning;

(3-2-2) inputting a data set B ^* having an input vector v into one of the decision trees of S-RF _t ;

(3-2-3) generating a split function f(v) at node O;

(3-2-4) calculating an information gain ΔE at node O;

(3-2-5) calculating cross-entropy Tr(Te, S) _t ; And

(3-2-6) If the calculated cross-entropy is less than the minimum threshold for stopping boosting, the current S-RF _t is stored, and otherwise, the step of re-performing from step (3-2-2) is T It is possible to learn the student random forest by executing until a random decision tree is constructed.

According to the teacher-student framework for lightening the ensemble classifier combined with the deep network and the random forest proposed in the present invention and a classification method based on the same, a new ensemble classifier is developed by combining a deep network and a random forest, By training the student model using the soft target data set B ^* , which is the output of the teacher model, the ensemble classifier developed through the teacher-student framework consisting of the teacher model and the student model is lightened and output more flexible classification results. The teacher model can be trained using the data set A including the class label and the data set B not including the class label, thereby preventing overfitting of the teacher model.

FIG. 1 is a diagram illustrating a teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention, and a classification method based thereon.
FIG. 2 is a diagram showing the configuration of a teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention.
3 is a diagram illustrating a flow of a classification method based on a teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention.
4 is a diagram illustrating a detailed flow of step S100 in a classification method based on a teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention.
5 is a diagram illustrating a detailed flow of step S300 in a classification method based on a teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention.
FIG. 6 is a diagram illustrating an algorithm for explaining a student network learning procedure in step S310 in a teacher-student framework-based classification method for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention. .
7 illustrates an algorithm for explaining a student random forest learning procedure in step S320 in a teacher-student framework-based classification method for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention. drawing.
8 is a diagram illustrating, for example, classifying a pedestrian direction class in a teacher-student framework for weight reduction of an ensemble classifier combined with a deep network and a random forest, and a classification method based thereon.
9 is a view comparing pedestrian direction estimation results of eight experiments including a classification method based on a teacher-student framework for lightening an ensemble classifier combined with a deep network and a random forest according to an embodiment of the present invention.
FIG. 10 is a diagram showing POE classification accuracy (Acc) for each direction class classified using a teacher-student framework for lightening an ensemble classifier combined with a deep network and a random forest according to an embodiment of the present invention as a confusion matrix .
11 is a diagram showing an experiment result for determining the number of trees of a student random forest in a teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention.
12 shows the accuracy of four experiments, the number of parameters, and the number of operations including a classification method based on a teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention. Compared drawings.
13 is a diagram summarizing experimental results for five CNN-based methods including a teacher-student framework-based classification method for weight reduction of an ensemble classifier combined with a deep network and a random forest according to an embodiment of the present invention. .
FIG. 14 shows the results of POE classification of (a) TUD and (b) KITTI data sets using a teacher-student framework for weight reduction of an ensemble classifier combined with a deep network and a random forest according to an embodiment of the present invention. One drawing.

Hereinafter, preferred embodiments will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art may easily implement the present invention. However, in describing a preferred embodiment of the present invention in detail, if it is determined that a detailed description of a related known function or configuration may unnecessarily obscure the subject matter of the present invention, the detailed description thereof will be omitted. In addition, the same reference numerals are used throughout the drawings for portions having similar functions and functions.

In addition, throughout the specification, when a part is said to be connected to another part, this includes not only the case that it is directly connected, but also the case that it is indirectly connected with another element interposed therebetween. In addition, the inclusion of certain components means that other components may be further included rather than excluding other components unless specifically stated to the contrary.

Deep learning networks require many parameters to create deep models. Therefore, a large amount of memory and time are required to perform a large amount of multiplication. In the present invention, in order to solve the shortcomings of the deep network model, a teacher-student framework was adopted, and a shallower student model having the same level of performance was constructed based on the deep teacher network.

FIG. 1 is a diagram illustrating a teacher-student framework for weight reduction of an ensemble classifier combined with a deep network and a random forest according to an embodiment of the present invention, and a classification method based thereon. As shown in FIG. 1, a teacher-student framework for weight reduction of an ensemble classifier combined with a deep network and a random forest according to an embodiment of the present invention and a classification method based on the same, include a deep network and a random forest. Combined to develop a new ensemble classifier, and train the student model with the soft target data set B ^* as an output of the teacher model, and use the ensemble classifier developed through a teacher-student framework consisting of a teacher model and a student model. It is possible to output more flexible classification results while being lighter, and by training the teacher model using dataset A with class labels and dataset B without class labels, overfitting of the teacher model can be prevented. I can.

As shown in Fig. 1, the teacher model generates a soft target (probability value) for each class by combining the output of the teacher deep network and the teacher random forest, and inputs the soft target value to train the student model. have. More specifically, referring to FIG. 1 in detail, (a) data set A labeled as a hard target is input to (b) deep teacher network and (c) teacher random forest, and (d) unlabeled data Set B can be entered into the two trained teacher models. (e) Combine the soft outputs of two teachers (teacher deep network and teacher random forest) into one soft target vector, (f) input the soft target data set B ^* into the student model, and (g) train the student model. So, (h) the final class probability can be obtained. As described above, the teacher-student framework of the present invention can reduce the size of the network and easily construct a compressed student model that can mimic the classification function of the teacher model.

FIG. 2 is a diagram illustrating a configuration of a teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention. As shown in FIG. 2, a teacher-student framework for weight reduction of an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention is, by using a data set A, a deep teacher network and a random teacher. A teacher learning module 100 that trains a teacher model composed of a forest, and a data set B is input to the deep teacher network and teacher random forest learned in the teacher learning module 100, and the two outputs are combined to provide soft target data. generating a soft target data to generate a set of B ^* module 200, a soft target data generated by using the data set B ^* generated by the module 200, the student study to study the student model consisting of student network and student Random Forest It may be configured to include a module 300 and a classification module 400 for performing classification by combining two outputs of the student network and the student random forest learned in the student learning module 300.

In addition, as shown in FIG. 2, the teacher learning module 100 includes a first teacher learning unit 110 for learning a deep teacher network using a data set A, and a feature map of the deep teacher network. It may be configured to include a second teacher learning unit 120 for learning the teacher random forest using, the student learning module 300, using the data set B ^* generated by the soft target data generation module 200 Thus, by using the first student learning unit 310 for learning the student network, and the data set B ^* generated by the soft target data generation module 200, the second student learning unit 320 for learning a student random forest It can be configured to include.

Each component of the teacher-student framework for weight reduction of the ensemble classifier in which the deep network and the random forest according to an embodiment of the present invention are combined will be described in detail in each step of the teacher-student framework-based classification method below. Let me explain.

3 is a diagram illustrating a flow of a classification method based on a teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention. As shown in FIG. 3, the classification method based on the teacher-student framework for lightening the weight of the ensemble classifier combined with the deep network and the random forest according to an embodiment of the present invention includes a teacher model using a data set A. Learning (S100), inputting data set B to the deep teacher network and teacher random forest, and combining the two outputs to generate soft target data set B ^* (S200), using data set B ^* It may be implemented including the step of training a student model (S300) and a step of performing classification (S400) by combining the two outputs of the learned student network and the student random forest, and preprocessing the input image by applying wavelet transformation. It may be implemented by further including the step (S10) of performing.

Hereinafter, each flow of a classification method based on a teacher-student framework for lightening the weight of an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention will be described in detail.

In step S100, a teacher model including a deep teacher network and a teacher random forest may be trained using the data set A. Step S100 may be processed by the teacher learning module 100.

The teacher model can be constructed using a deep teacher network and a teacher random forest (RF) with a high level of performance based on a large amount of learning data. Here, the data set A may be a hard target data set including a class label. That is, in step S100, the teacher model may be trained using the data set A labeled 0 or 1.

The deep teacher network may apply a softened softmax function to obtain a soft target output that is a probability value of each class. That is, different from the softmax function of the general CNN model, the deep teacher network T is a softened softmax function as shown in Equation 1 below to obtain a soft target (output probability). softmax function) can be applied to the vector of the teacher pre-softmax activations aT. The basic idea of the teacher deep network is to enable the student network to capture not only the information provided by the actual label, but also the smaller structures learned by the teacher deep network.

Here, aT is very close to one hard target expression for the true label of the sample, but the softened output (P _T ) of the teacher can be distributed more smoothly as the temperature (τ>1) increases.

This method can smooth the signals coming out of the teacher deep network's output and provide more information to the student network while learning the student model. However, since the student network's performance is sensitive to temperature, this value must be determined empirically for all training data, and considerable effort is required to predict the optimal temperature.

In step S100 of the present invention, in order to reduce the effort required to determine the temperature and obtain the soft output of the teacher model, as shown in FIG. 1, a new soft output is selected by combining the soft output of the teacher deep network and the teacher random forest. I did. Random Forest, which is a decision tree ensemble classifier, is known to process a very large amount of data at a higher learning speed than conventional classifiers. Also, random forests essentially provide a smoother distribution of classification results for a particular class.

4 is a diagram illustrating a detailed flow of step S100 in a classification method based on a teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention. As shown in FIG. 4, step S100 of a classification method based on a teacher-student framework for lightening an ensemble classifier combined with a deep network and a random forest according to an embodiment of the present invention is performed using a data set A. It may be implemented including the step of learning the deep teacher network (S110) and the step of learning the teacher random forest using the feature map of the deep teacher network (S120).

In step S110, the deep teacher network may be trained using the data set A. Step S110 may be processed by the first teacher learning unit 110. More specifically, in step S110, the deep teacher network is first trained using the training data set A, in which case the data set A may be a hard target data set including a class label. Data set A={(x _i , y _i )|i=1, 2,… , N} is the M-dimensional input vector _{x i = (x i1, x} i2, ..., x iM) and scalar class label is a display of the expert _{x i y i = {g 1} , g 2, ... , g _c }.

The deep teacher network is based on the ResNet-101 model (He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition, In Proceedings of IEEE Conference of Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV USA, 26 June 1 July 2016; pp. 770-778.) can be generated as class-labeled (hard target) training data. The structure of the deep teacher network may consist of 101 parameter layers, one average pooling layer, and one fully connected layer. ResNet adds a shortcut connection that skips one or more layers to each 33 filter pair, but the basic architecture can be the same as a normal CNN. In addition, ResNet can increase the number of dimensions by using identity mapping for all shortcuts and zero-padding. The output of the short link can be added to the output of the stacked layer.

Given the training dataset A, we can fine-tune the pretrained ResNet-101 model on ImageNet to a smaller dataset A, updating all network weights for the new task. After learning the deep teacher network, Equation 1 may provide a soft output to the output unit (class). That is, the output probability may be provided to classes classified by a predetermined number.

In step S120, a teacher random forest may be trained using a feature map of the deep teacher network. Step S120 may be processed by the second teacher learning unit 120. That is, in step S120, the individual decision tree of the teacher random forest as the second classifier may be learned using the final feature vector for the input vector x _i with the hard class label y _i . The learning of the decision tree is based on random sampling of subsets and selection of the separation function using information gain. The final class distribution of sample x may be generated using an ensemble (arithmetic mean) of each class probability distribution p _t (c _i |x) of all trees T, as shown in Equation 2 below.

In step S200, the data set B is input to the teacher model learned in step S100, and the soft target data set B ^* may be generated using the output soft output. Step S200 may be processed by the soft target data generation module 200. More specifically, in step S200, as shown in (e) of FIG. 1, data set B is input to the deep teacher network and teacher random forest learned in step S100, and the output of the teacher deep network and the teacher random forest is output. By combining into one soft target vector, a soft target data set B ^* , which is a probability value of each class, can be generated. In this case, in step S200, the data set B may be a data set that does not include a class label.

More specifically, after training of the teacher model is complete, a much larger, unlabeled training data set B is applied to the teacher model, and a new data set B ^* consisting of a soft target (class probability) as opposed to a hard target is constructed. I can.

Unlike the algorithm that generates a soft target by applying only one data set A to the teacher model, the present approach can use an additional training data set B to prevent overfitting of the teacher model. The new soft target data set B ^* generated in step S200 may capture more information than the original hard target data by maintaining a relationship between different classes. In addition, it is possible to obtain a more flexible classification result than using a hard target data set.

After all M samples x in data set B have been trained, the new data set B ^* expressed as class probability p _i ^* (soft target) is B ^* ={(x _i , p _i ^* )|i=1 , 2, … , M} can be created. The difference in recognition performance may occur depending on the number of decision trees used in the random forest.

Meanwhile, a classification method based on a teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention may be implemented by further including step S10. That is, in step S10, the preprocessing module 500 may perform preprocessing on the input image by applying the wavelet transform. In particular, step S10 may be additionally processed when classifying an image by applying a classification method based on a teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest according to an embodiment of the present invention are combined. have. More specifically, in step S10, two high-pass filtered sub-images and one low-pass filtered sub-image are generated, and in step S100, the three sub-images generated in step S10 are generated. Teacher models can be trained using images.

Specifically, in addition to the soften softmax function, three manual filter responses of wavelet transform can be provided to the model. That is, using two high-pass filtered sub-images (LH and HL) and one low-pass filtered sub-image LL, ( Using the Daubechies D4 wavelet with the gray image as shown in a) to provide appropriate manual characteristics can improve the results for specific classification problems. Further, since the wavelet transform has good spatial frequency region characteristics and can preserve spatial information and gradient information of an image, it may be helpful to improve classification performance under various brightness conditions.

In step S300, a student model including a student network and a student random forest may be trained using the data set B ^* generated in step S200. Step S300 may be processed by the student learning module 300.

After training the teacher model, a lightweight student model can be constructed using the soft target data set B ^* generated from the teacher model. As in the teacher model, the student model can consist of 1 student network and 1 student random forest.

5 is a diagram illustrating a detailed flow of step S300 in a classification method based on a teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention. As shown in FIG. 5, step S300 of a classification method based on a teacher-student framework for lightening an ensemble classifier combined with a deep network and a random forest according to an embodiment of the present invention uses a data set B ^* Thus, it may be implemented including the step of learning the student network (S310) and the step of learning the student random forest using the data set B ^* (S320).

In step S310, the student network may be trained using the data set B ^* generated in step S200. Step S310 may be processed by the first student learning unit 310.

Student networks can be created by modifying the Darknet reference model. Available online: https://pjreddie.com/darknet/imagenet/#reference (accessed on 27 December 2018). This is because the calculation speed of the DarkNet reference model is 16 times faster than the existing ResNet-101 and 2 times faster than AlexNet in one CPU when the number of parameters is 1/5 and 1/10. So, instead of compressing the deep teacher network, we can use the shallow DarkNet reference model as the student network and retrain the student network using the soft target data set generated from the teacher model.

The structure of the student network may consist of a total of 8 convolutional layers including 7 max pooling layers, and one average pooling layer after each convolution layer. The front seven convolution layers have a 3×3 convolution filter and a max pooling layer with a 2×2 size filter, and the last convolution layer has a 1×1 size convolution filter, and a fully connected layer. There is an average pooling layer instead of a (fully connected layer), which avoids over-fitting problems and reduces the number of learnable parameters of the fully connected layer. In addition, batch normalization is applied to each convolutional layer, and leaky ReLU (LReLU) can be used as an activation function to solve the dying ReLU problem. The LReLU function f(x) has a small negative value when x<0 instead of 0 as shown in Equation 3 below.

To train the student network, we can use pretrained convolutional weights in ImageNet and perform fine tuning using the soft target data set B ^* . The cross-entropy criterion may be based on frame-by-frame minimization by replacing the hard target vector with a soft target vector as shown in Equation 4 below.

Where N is the number of samples in the data set B ^* and C is the number of classes. Also, P _T (x _i |c _j ) and P _S (x _i |c _j ) are the posterior class probabilities of the teacher and student for the input vector x _i , respectively.

FIG. 6 is a diagram illustrating an algorithm for explaining a student network learning procedure in step S310 in a teacher-student framework-based classification method for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention. to be. In step S310, the first student learning unit 310 may first learn the student network with an algorithm as shown in FIG. 6.

를 W_S ^*로 업데이트하는 단계, 및 (3-1-5) 학생 네트워크를 위한 최적 파라미터 W_S ^*를 선택하는 단계를 수행하여, 학생 네트워크를 학습시킬 수 있다.
More specifically, in step S310, (3-1-1) initializing the parameter W _S of the student network, (3-1-2) inputting the data set B ^* into the pre-trained network, (3-1-3) Calculating the loss function (L(Ws)), (3-1-4)

By performing the steps of updating to W _S ^* and (3-1-5) selecting the optimal parameter W _S ^* for the student network, the student network may be trained.

In step S320, the student random forest may be trained using the data set B ^* generated in step S200. Step S320 may be processed by the second student learning unit 320.

Similar to the student network, the initial decision tree of the student random forest can use the same structure as the teacher random forest, such as a split function with the number of trees, the tree depth, and a split threshold for each node of an individual tree.

The learning of the decision tree of the student random forest takes an input vector v and a class probability p _i ^* as inputs. The input vector has 128 dimensions created from the last feature maps (4×4×8), and the class probability can be estimated from the output of the teacher model. From the training data composed of the output vector with the class probabilities of the soft target data set B ^*, the decision tree of the student random forest can randomly select p variables with the class probabilities from the samples.

B _'O Let denote a sample from node O. Trained in advance and a randomly generated split function _{(split function) f (v p} ) is repeatedly, in the node O left (B 'random subset (subset) B _O to _l) and right (B' _r) subset Can be divided. To select the best split function, the entropy E(O) of node O can be estimated using only the p variable with probability distribution P _j ^* . In the present invention, the entropy E (B' _O ) of the node O may be defined as in Equation 5 below.

Using the same method, the left and right sub-set of the node O is divided into B _'l and B' _r, is the entropy E (B _'l) and E (B' _r) it can be calculated. From the three entropies, the information gain ΔE of node O can be calculated from Equation 6 below.

This process can be repeated while applying the number of candidate split functions, and the function with the maximum ΔE can be determined as the best split function f(v _p ) for node O.

After the initial decision tree Tr _t is expanded, the probability distribution of C classes may be stored in a leaf node. Then, P _ij ^* (Te) representing the j-th class distribution of sample i of the data set B ^* transcribed by the teacher random forest in Equation 6, and a sample based on the constructed decision tree t With P _ij ^* (S _t ) representing the j-th class distribution of i, cross-entropy can be estimated. The general form of the final cross-entropy is shown in Equation 7 below.

Thanks to the high performance of the boosted random forest, boosting can be repeated to update the t-th weak decision tree until Tr(Te, S) _t falls below the minimum criterion. When T random decision trees are completed, the student random forest may finally become a T tree composed of probability distributions per class.

7 illustrates an algorithm for explaining a student random forest learning procedure in step S320 in a teacher-student framework-based classification method for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention. It is a drawing. In step S320, the second student learning unit 320 may learn the student random forest using an algorithm as shown in FIG. 7.

More specifically, in step S320, (3-2-1) the t-th tree structure (T-RF _t ) of the teacher random forest is copied to the _t -th tree structure (S-RF _t ) of the student random forest, and prior learning (Transfer Learning), (3-2-2) inputting the data set B ^* with the input vector v into one of the decision trees of S-RF _t , (3-2-3) at node O Generating split function f(v), (3-2-4) calculating information gain ΔE at node O, (3-2-5) cross-entropy Tr( Te, S) calculating _t , and (3-2-6) if the calculated cross-entropy is less than the minimum threshold θ to stop boosting, store the current S-RF _t , otherwise step (3- The re-performing steps from 2-2) are performed until T random decision trees are constructed, so that the student random forest can be learned.

In step S400, classification may be performed by combining two outputs of the student network and the student random forest learned in step S300. Step S400 may be processed by the classification module 400. More specifically, in step S400, a final probability is generated by combining the output values of the student network and the student random forest learned in step S300, and through this, the final probability may be classified as a class having the highest final probability. That is, in the present invention, by inputting the learned student model and combining the output value of the student network and the output value of the student random forest to generate a final probability, it is possible to more accurately classify.

Experiment result

In order to demonstrate the efficiency of the teacher-student framework and classification method based on the teacher-student framework for lightening the ensemble classifier combined with the deep network and the random forest according to an embodiment of the present invention, conflict in ADAS (Advanced Driver Assistant System) The performance of the present invention was evaluated by applying it to the estimation of a pedestrian's posture direction for avoidance, and a comparative experiment was performed using another approach suggested in a recent study. In ADAS, a vehicle can detect a pedestrian and predict a pedestrian's intention in advance based on the pedestrian's Pose Orientation Estimation (POE). Therefore, when a pedestrian is stepping on the road without noticing the vehicle, it is possible to warn the driver, thereby greatly reducing the possibility of a collision. At this time, since POE in a single image captured by a moving vehicle is aimed at, 3D POE using a stereo camera or an RGBD sensor is not considered.

The performance of the present invention was evaluated using a benchmark database, and comparative experiments were conducted using different approaches presented in recent studies.

In this experiment, first, to prove that the teacher-student framework for lightening the ensemble classifier combined with the deep network and the random forest, and the classification method based on it, is effective in estimating the direction of various pedestrian postures, the performance of POE is verified. I did. The experiment was conducted using an Intel Core i7 processor with 24GB of RAM running Microsoft Windows 10. In addition, all RF approaches, including teacher random forest and student random forest, were implemented based on CPU, and the teacher deep network was implemented using a single Titan Xp GPU.

For learning of the deep teacher network, the batch size, momentum, learning rate, and weight decay were set to 32, 0.9, 0.001, and 0.0005, respectively. In the case of a random forest, an important parameter in terms of performance and memory required to store the tree is the depth of the tree and its number. In this experiment, the maximum tree depth was set to 20 and the number of trees in the teacher's random forest was set to 300. To determine the number of trees in the student random forest, the number of trees was sequentially reduced to 250, 200, 150, 100, 70, and 50, and for more accurate and faster calculations, as described later based on the experimental results. It was set to 70.

The student network and the student random forest were retrained using a soft target learning data set B ^* based on Algorithm 1 and Algorithm 2 as shown in FIGS. 6 and 7. Although many data sets related to pedestrian detection are available, relatively few studies have performed pedestrian direction estimation. Therefore, in the present invention, a POE experiment was performed using the most popular TUD multi-view pedestrian data set consisting of 5,228 pedestrian images with a bounding box and discrete direction annotation. This data set contains 4,732 full-body pedestrian images for training, 248 validation tests, and 248 tests. The images in the TUD data set were taken in real street situations, and every image included a variety of poses and clothes, making the data set even more challenging. It is a common fact that a model trained on a small data set does not generalize the data in the validation and test set, resulting in overfitting results. To reduce overfitting, the present invention increases the size of the data set by applying data magnification such as image movement, enlargement and reduction, rotation at any angle between -15 degrees and +15 degrees, left-right flip and cropping. Made it. All training images include the original and duplicate images provided in the teacher model. By the aforementioned data increase, 4,732 images were allocated to data set A and 4,732 images were allocated to data set B.

FIG. 8 is a diagram illustrating an example of classifying a pedestrian direction class in a teacher-student framework for lightening an ensemble classifier combined with a deep network and a random forest, and a classification method based thereon. In pedestrian direction estimation, the number of classes may increase when all directions are predicted. Therefore, most of the existing studies have used a method of dividing the direction into N clusters and recognizing them as shown in FIG. For example, the TUD data set consists of 5,228 pedestrian images with directional annotations such as behind, front, left, right, left rear, right rear, left front, and right front of a bounding box. For the TUD data set, the orientation class is divided into 45 degrees.

In order to verify the validity of the direction estimation of the eight classes, the precision, recall and false positive rate (FPR) of the TUD data set were measured. This value is generally used to evaluate object recognition performance. Also, accuracy (Accuracy) is used to compare performance between classes by evaluating poses and confusion matrices. Accuracy is the ratio of successful detection to the total number of cases investigated.

Performance evaluation on TUD data set

In order to verify the effect of POE classification using the present invention, five state-of-the-art methods and performance were compared. Each experiment is as follows. (1) MoAWG classifying POE using an array of highly randomized tree classifiers, (2) PLS-RF using partial least squares-based model combined with random forest classifier, (3) recognizing body posture orientation MACF using a sparse representation technique, (4) VGG-16 using 16 weighted CNN layers and CNN with low-resolution images, (5) ResNet-101 based on deep residual nets, ( 6) Proposed teacher model without manual filter, (7) Proposed T-Model, (8) Proposed S-Model including student network and student random forest. Of the total eight methods, methods (4) to (8) are based on CNN.

9 is a view comparing pedestrian direction estimation results of eight experiments including a classification method based on a teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention. . 9 compares the results of eight approaches in terms of average precision (AP), average recall (AR), and average FPR (AFPR). As shown in FIG. 9, in all experiments, it can be seen that the CNN-based method (methods (4) to (8)) has better classification performance than the conventional manual and classifier-based methods. MoAWG achieved the highest performance among the three existing approaches (MoAWG, PLS-RF, MACF), but 0.2%, 3.2% and 0.6% lower performance than VGG-16, the lowest performance among deep network-based approaches. Showed. VGG-16 and ResNet-101 showed better performance than the conventional approach, but since they use the basic CNN model, it can be confirmed that their performance is lower than that of the present invention. Among the three proposed methods, the T-Model method showed the best performance compared to other methods in the three evaluation items applied because it uses the deep teacher network and the teacher random forest at the same time.

In the case of the proposed T-Model without handcraft filters (T-Model), which does not have a pre-processing step S10 using a manual filter, performance was poor in all three evaluation items compared to the original T-Model. Based on the results, it can be seen that the wavelet transform has good spatial frequency position characteristics and can preserve spatial information and longitude information of an image.

The evaluation result of the proposed S-Model showed slightly lower performance of 7.3% and 5.3% in terms of AP and AR compared to the T-Model. However, since there is less performance degradation compared to the size reduction ratio of the model, it can be seen that the proposed method effectively improves the memory and speed requirements while maintaining the performance. Compared with other CNN-based methods, the proposed S-Model has relatively high AP and AR and low AFPR. This indicates that the proposed method is robust against complex backgrounds or blurry pedestrian outlines.

FIG. 10 is a diagram showing POE classification accuracy (Acc) for each direction class classified using a teacher-student framework for lightening an ensemble classifier combined with a deep network and a random forest according to an embodiment of the present invention as a confusion matrix to be. As shown in FIG. 10, most directions except for'Back'and'Lback' showed similar classification performance. The main reason for the low accuracy of these two directions compared to the other directions is that even though the wavelet transform was applied in the previous step of the CNN, the two directions had similar shapes. On the other hand,'Lfront'and'Rback' showed the best classification performance due to the difference in appearance.

Determining the optimal number of decision trees for student RF

In the case of a student random forest, the number of decision trees is an important factor in reducing the number of parameters for saving processing time and memory. The number of trees was sequentially reduced to 200, 150, 100, 70, and 50, comparing the precision, recall, and accuracy performance on the TUD data set to determine the optimal number of trees for the student random forest. The experiment was carried out.

11 is a diagram showing an experiment result for determining the number of trees of a student random forest in a teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention. As shown in FIG. 11, as the number of trees increases, the precision, the recovery rate, and the accuracy increase, but the number of parameters relatively increases and the speed and compression rate decrease. Based on these results, 70 trees can be considered as the optimal number of trees in the student random forest because they show similar or slightly higher performance than other trees. Therefore, in the present invention, the number of trees in the student random forest is set to 70 in order to increase the accuracy and reduce the number of parameters.

Model compression evaluation

The goal of model compression is to create an optimal student model with fewer parameters and operations with similar performance to the teacher model. Therefore, the proposed student model was compared in terms of the number of parameters and computation using the popular model compression method, MobileNet and TUD data set. The comparative model was fine-tuned using TUD training data based on pretrained parameters. MobileNets are based on separable depth-wise convolutions that are applied to reduce the number and operation of parameters. In this experiment, three comparison methods were performed using one Titan-X GPU.

12 shows the accuracy of four experiments, the number of parameters, and the number of operations including a classification method based on a teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest are combined according to an embodiment of the present invention. It is a comparison drawing. As shown in FIG. 12, the proposed student model can reduce the number of parameters by about 5 times and the number of operations by about 19.6 times compared to the teacher model. That is, it can be seen that the POE classification accuracy of the student model is somewhat lower than that of the teacher model, but the number of required operations and parameters is very small. In addition, the student model has 17.9% better POE accuracy than MobileNet, and uses 5 times fewer operations. However, the student network uses the general convolution method, which increases the number of operations by 19.6 times. As can be seen from the comparison result, it can be seen that the proposed model compression method shows superior performance in terms of POE classification accuracy and number of operations compared to the conventional compression method.

Performance evaluation on the KITTI data set

In order to verify whether the used algorithm proposed in the present invention can be effectively applied to other data sets, the algorithm of the present invention was applied to the KITTI data set and the results were compared.

The KITTI data set used as the second data set is a real world computer including stereo imaging, optical flow, visual odometry, 3D object detection, and 3D tracking. It is a vision benchmark. Among the nine categories available, an experiment was performed on the pedestrian category. The pedestrian category of the KITTI data set was divided into a training data set composed of 5,415 images and a validation set composed of 2,065 images. In addition, we applied data increase to only the training data set to increase the size of the data set and used the entire training data set of 4,732 images. The difficulty of the data set was defined as “easy”, “moderate”, and “hard” according to size, occlusions and truncation level. Detection of an insignificant area or a detection smaller than the minimum size is not regarded as a false positive. To train the student model on the KITTI data set, the training data was applied to the teacher model, and the soft target data, the output of the teacher model, was applied to the student network and the student random forest. While the model was being trained, the pedestrian's direction was normalized at eight angles, and successive direction values were estimated using Equation 8.

In order to verify the validity of the direction estimation of eight classes, in the case of the KITTI data set, since the pedestrian data of the KITTI data set is continuously displayed in a different direction than the TUD, Average Orientation Similarity (AOS) was used. .

To evaluate the performance, accuracy was compared with the following state-of-the-art methods. (1) DPM-VOC+VP, which deals with different perspectives by expanding the deformable part of the model method, (2) Mono3D, which detects 3D objects from single monocular images using CNN, (3) subcategories SubCNN based on cognitive convolutional neural network, (4) FRCNN using viewpoint inference from the top of highly optimized CNN-based detection framework, (5) Proposed student model.

13 is a diagram summarizing experimental results for five CNN-based methods including a teacher-student framework-based classification method for weight reduction of an ensemble classifier combined with a deep network and a random forest according to an embodiment of the present invention. to be. As shown in FIG. 13, in the experiment using the KITTI data set, the two methods, DPM-VOC+VP and FRCNN, showed a lower AOS ratio than the other three methods when the input image was small due to the blurring of the pedestrian. However, since the SubCNN method uses an image pyramid to detect small pedestrians, AOS performance is better than other methods. SubCNN showed a relatively excellent AOS ratio with respect to the KITTI data set, but since the network structure is deeper and wider than the student model proposed in the present invention, there is a disadvantage in that an additional network for region proposal and object detection is required. However, the teacher-student framework for weight reduction of an ensemble classifier in which a deep network and a random forest according to an embodiment of the present invention are combined to improve AOS speed by applying a teacher-student structure, and two compressed classifiers (student Network and Student Random Forest) compensated for the shortcomings of the others, and showed excellent performance for the easy, normal and difficult data of the KITTI data set.

FIG. 14 shows the results of POE classification of (a) TUD and (b) KITTI data sets using a teacher-student framework for weight reduction of an ensemble classifier combined with a deep network and a random forest according to an embodiment of the present invention. It is a drawing. As shown in Fig. 14, the student model proposed in the present invention can correctly predict the pedestrian's direction even when the pedestrian's body is distorted or partially obscured by another pedestrian, even if the image is blurred.

As described above, in the present invention, the teacher model may generate a probability value for each class by combining the output of the deep teacher network and the output of the teacher random forest, and input the soft target value to train the student model. By combining two different classification models, not only can the model size be reduced, but a student model can be constructed that mimics the classification function of the teacher model. In addition, unlike the existing CNN-based classification approach, the present invention selects a new soft output by combining the output of the teacher deep network and the teacher random forest, and constructs a student model of equivalent performance based on the teacher model.

The present invention described above can be modified or applied in various ways by those of ordinary skill in the technical field to which the present invention belongs, and the scope of the technical idea according to the present invention should be determined by the following claims.

100: Teacher Learning Module
110: First Teacher Learning Department
120: Second Teacher Learning Department
200: soft target data generation module
300: Student Learning Module
310: 1st Student Learning Department
320: Second Student Learning Department
400: classification module
500: pretreatment module
S10: Step of performing preprocessing on the input image by applying wavelet transform
S100: training a teacher model using data set A
S110: Training the deep teacher network using the data set A
S120: Learning a teacher random forest using the feature map of the deep teacher network
S200: Inputting the data set B into the deep teacher network and the teacher random forest, and combining the output two outputs to generate a soft target data set B ^*
S300: training a student model using the data set B ^*
S310: Learning a student network using the data set B ^*
S320: Learning a student random forest using the data set B ^*
S400: combining the two outputs of the learned student network and the student random forest to perform classification

Claims (20)

As a teacher-student framework,
A teacher learning module 100 for learning a teacher model composed of a teacher deep network and a teacher random forest using the data set A;
A soft target data generation module 200 for inputting data set B into the deep teacher network and teacher random forest learned in the teacher learning module 100, and combining the output two outputs to generate a soft target data set B ^* ;
A student learning module 300 for learning a student model composed of a student network and a student random forest by using the data set B ^* generated by the soft target data generation module 200; And
The ensemble classifier combined with the deep network and the random forest, characterized in that it comprises a classification module 400 for performing classification by combining two outputs of the student network and the student random forest learned in the student learning module 300 Teacher-student framework for lightweighting. The method of claim 1, wherein the data set A,
A teacher-student framework for lightening an ensemble classifier combined with a deep network and a random forest, characterized in that it is a hard target data set including class labels. The method of claim 1, wherein the data set B,
A teacher-student framework for lightweighting an ensemble classifier combining a deep network and a random forest, which is a data set that does not contain a class label. The method of claim 1, wherein the teacher learning module (100),
A first teacher learning unit 110 for learning the deep teacher network using the data set A; And
And a second teacher learning unit 120 that learns a teacher random forest using the feature map of the deep teacher network.For lightening the ensemble classifier in which the deep network and the random forest are combined Teacher-student framework. The method of claim 1, wherein the deep teacher network,
A teacher-student framework for weight reduction of an ensemble classifier combined with a deep network and a random forest, characterized by applying a softened softmax function to obtain a soft target output, which is a probability value of each class. The method of claim 1,
A teacher-student framework for lightening an ensemble classifier in which a deep network and a random forest are combined, characterized in that it further comprises a preprocessing module 500 that performs preprocessing on the input image by applying wavelet transform. The method of claim 1, wherein the student learning module (300),
A first student learning unit 310 for learning a student network by using the data set B ^* generated by the soft target data generation module 200; And
Using the data set B ^* generated by the soft target data generation module 200, characterized in that it comprises a second student learning unit 320 for learning a student random forest, a deep network and a random forest are combined. A teacher-student framework for lightweight ensemble classifiers. The method of claim 7, wherein the first student learning unit 310,
(3-1-1) initializing the parameter (W _S ) of the student network;
(3-1-2) inputting the data set B ^* into a pre-trained network;
(3-1-3) calculating a loss function (L(Ws));
(3-1-4)

Updating to W _S ^* ; And
(3-1-5) A teacher for lightening an ensemble classifier combined with a deep network and a random forest, characterized in that the student network is trained by performing the step of selecting the optimal parameter W _S ^* for the student network -Student framework. The method of claim 8, wherein in the step (3-1-3),
A teacher-student framework for lightweighting an ensemble classifier combined with a deep network and a random forest, characterized in that the loss function is calculated using the following equation.

In the above equation, N is the number of samples in the data set B ^* , C is the number of classes, and P _T (x _i |c _j ) and P _S (x _i |c _j ) are the teachers and students for the input vector x _i , respectively. This is the posterior class probability. The method of claim 7, wherein the second student learning unit (320),
(3-2-1) copying the t-th tree structure (T-RF _t ) of the teacher random forest to the t-th tree structure (S-RF _t ) of the student random forest to perform transfer learning;
(3-2-2) inputting a data set B ^* having an input vector v into one of the decision trees of S-RF _t ;
(3-2-3) generating a split function f(v) at node O;
(3-2-4) calculating an information gain ΔE at node O;
(3-2-5) calculating cross-entropy Tr(Te, S) _t ; And
(3-2-6) If the calculated cross-entropy is less than the minimum threshold for stopping boosting, the current S-RF _t is stored, and otherwise, the step of re-performing from step (3-2-2) is T A teacher-student framework for lightening an ensemble classifier combining a deep network and a random forest, characterized in that the student random forest is trained by performing until constructing two random decision trees. As a classification method based on the teacher-student framework,
(1) using the data set A, training a teacher model composed of a deep teacher network and a teacher random forest;
(2) inputting the data set B into the deep teacher network and the teacher random forest learned in step (1), and combining the two outputs to generate a soft target data set B ^* ;
(3) learning a student model consisting of a student network and a student random forest by using the data set B ^* generated in step (2); And
(4) For reducing the weight of the ensemble classifier in which the deep network and the random forest are combined, comprising the step of performing classification by combining the two outputs of the student network and the student random forest learned in step (3). Classification method based on the teacher-student framework. The method of claim 11, wherein the data set A,
A classification method based on a teacher-student framework for lightening an ensemble classifier combined with a deep network and a random forest, characterized in that it is a hard target data set including a class label. The method of claim 11, wherein the data set B,
Classification method based on a teacher-student framework for weight reduction of an ensemble classifier combined with a deep network and a random forest, characterized in that the data set does not include a class label. The method of claim 11, wherein the step (1),
(1-1) learning the deep teacher network using the data set A; And
(1-2) A teacher for weight reduction of an ensemble classifier combining a deep network and a random forest, comprising the step of learning a teacher random forest using a feature map of the deep teacher network- Classification method based on student framework. The method of claim 11, wherein the deep teacher network,
Based on a teacher-student framework for weight reduction of an ensemble classifier combined with a deep network and a random forest, characterized by applying a softened softmax function to obtain a soft target output, which is a probability value of each class Classification method. The method of claim 11,
(0) A classification method based on a teacher-student framework for lightening an ensemble classifier combined with a deep network and a random forest, further comprising the step of performing pre-processing on the input image by applying a wavelet transform. The method of claim 11, wherein the step (3),
(3-1) learning a student network by using the data set B ^* generated in step (2); And
(3-2) For lightening the ensemble classifier in which the deep network and the random forest are combined, comprising the step of learning a student random forest using the data set B ^* generated in step (2). Classification method based on the teacher-student framework. The method of claim 17, wherein the step (3-1),
(3-1-1) initializing the parameter (W _S ) of the student network;
(3-1-2) inputting the data set B ^* into a pre-trained network;
(3-1-3) calculating a loss function (L(Ws));
(3-1-4)

Updating to W _S ^* ; And
(3-1-5) A teacher for lightening an ensemble classifier combined with a deep network and a random forest, characterized in that the student network is trained by performing the step of selecting the optimal parameter W _S ^* for the student network -Classification method based on student framework. The method of claim 18, wherein in the step (3-1-3),
A classification method based on a teacher-student framework for weight reduction of an ensemble classifier combined with a deep network and a random forest, characterized in that the loss function is calculated using the following equation.

In the above equation, N is the number of samples in the data set B ^* , C is the number of classes, and P _T (x _i |c _j ) and P _S (x _i |c _j ) are the teachers and students for the input vector x _i , respectively. This is the posterior class probability. The method of claim 17, wherein the step (3-2),
(3-2-1) copying the t-th tree structure (T-RF _t ) of the teacher random forest to the t-th tree structure (S-RF _t ) of the student random forest to perform transfer learning;
(3-2-2) inputting a data set B ^* having an input vector v into one of the decision trees of S-RF _t ;
(3-2-3) generating a split function f(v) at node O;
(3-2-4) calculating an information gain ΔE at node O;
(3-2-5) calculating cross-entropy Tr(Te, S) _t ; And
(3-2-6) If the calculated cross-entropy is less than the minimum threshold for stopping boosting, the current S-RF _t is stored, and otherwise, the step of re-performing from step (3-2-2) is T A classification method based on a teacher-student framework for lightening an ensemble classifier combined with a deep network and a random forest, characterized in that the student random forest is learned by performing until constructing two random decision trees.

Images