KR102049331B1 - Apparatus and method for classifying images, and apparatus for training images for classification of images - Google Patents

Abstract

The present invention classifies images by using the error function generated based on information on features included in an image and information on features included in a class in the image, and classifies the images based on the results. An apparatus and method are proposed. The apparatus according to the present invention generates an error function based on the information about the first features included in the reference image and the information about the second features located in the class formed in units of objects in the reference image, and based on the error function. An image learner configured to learn the input images; A threshold setting unit that sets a threshold based on a result obtained by learning the images; And an image classifier that classifies the images based on the threshold value.

Description

Apparatus and method for classifying images, and apparatus for training images for classification of images}

The present invention relates to an image learning apparatus for training neural networks for image classification. The present invention also relates to an image classification apparatus and method for classifying images based on image learning results.

In general, a convolutional neural network (CNN) is trained using a method of minimizing softmax errors. However, in the face recognition problem, the model trained using only the softmax error has difficulty. This is because the features derived from the models are concentrated together, making it difficult to distinguish between classes.

In order to classify images with high similarity, such as faces, not only the features to be learned must be classified, but also a model in which the features are distinguished and discriminated, that is, discriminated, is required.

Korean Laid-Open Patent No. 2014-0096595 (Published: 2014.08.06.)

The present invention has been made to solve the above-described problem, the image is trained using an error function generated based on the information about the features included in the image and the information contained in the class in the image, An object of the present invention is to propose an image classification apparatus and method for classifying images based on the results.

Another object of the present invention is to propose an image learning apparatus for learning images using an error function generated based on information on features included in an image and information on features included in a class in the image.

However, the object of the present invention is not limited to the above-mentioned matters, and other objects not mentioned will be clearly understood by those skilled in the art from the following description.

The present invention has been made to achieve the above object, the information on the first features included in the reference image and the second feature located in the class (object) formed in the unit (object) unit in the reference image An image learning unit configured to generate an error function based on the information and to learn the input images based on the error function; A threshold setting unit that sets a threshold based on a result obtained by learning the images; And an image classification unit for classifying the images based on the threshold value.

Preferably, the image learner learns the images using a convolutional neural network.

Advantageously, the image learner comprises a first error generated based on the average value of the first features, a second error generated based on the average value of the second features, and a third error associated with cross-entropy. Generate the error function based on.

Preferably, the image learner calculates the first error based on first difference values between the average value of the first features and each feature included in the reference image, and a first weight. More preferably, the image learner calculates a first value by applying a maximum value selected from the first difference values to L2-norm, and applies each first difference value to L2-norm to apply second values. The first error is calculated based on the results obtained by comparing the first values with zeros and the third values obtained by subtracting each second value and zero.

Preferably, the image learner may generate the second error based on the second difference values between the average value of the second features and each feature included in the reference image, and the number of classes included in the reference image. Calculate. More preferably, the image learning unit calculates second difference values for each class, calculates fourth values for each class by applying the second difference values calculated for each class to L2-norm, and calculates each class. The second error is calculated based on the result obtained by summing the fourth values.

Preferably, the image learner calculates a fifth value by multiplying the first error by a second weight, multiplies the second error by a third weight, and calculates a sixth value, and the fifth value and the sixth value. The error function is generated based on a result obtained by summing a value and the third error.

Advantageously, the image classifier classifies the images based on whether the features included in the images are below the threshold.

Preferably, when it is determined that the features included in the comparison target image are less than or equal to the threshold value, the image classification unit classifies the object included in the comparison target image into the same person as the object included in the reference image. If it is determined that the included features exceed the threshold, the objects included in the comparison target image are classified into objects and others included in the reference image.

The image classification apparatus may further include an image adjusting unit configured to edit each image to have a predetermined size based on a point where the first features are expected to be concentrated when the images are input. Part learns the edited images.

In addition, the present invention generates an error function based on the information about the first features included in the reference image and the information about the second features located in the class (object) formed in the object unit in the reference image Learning the input images based on the error function; Setting a threshold based on a result obtained by learning the images; And classifying the images based on the threshold value.

Advantageously, said training step trains said images using a convolutional neural network.

Advantageously, said learning step comprises a first error generated based on an average value of said first features, a second error generated based on an average value of said second features and a third associated with cross-entropy. The error function is generated based on the error.

Preferably, the learning step calculates the first error based on first difference values between the average value of the first features and each feature included in the reference image, and a first weight. More preferably, the learning may include applying a maximum value selected from the first difference values to L2-norm to calculate a first value, and applying each first difference value to L2-norm to a second value. The first error is calculated based on the results obtained by comparing the first values with zeros and the third values obtained by subtracting each second value and zero.

Preferably, the learning may include the second error value based on the second difference between the average value of the second features and each feature included in the reference image, and the number of the classes included in the reference image. To calculate. More preferably, the learning may be performed by calculating second difference values for each class, calculating second values by class by applying the second difference values calculated for each class to L2-norm, and calculating each class. The second error is calculated based on the result obtained by summing the fourth values.

Preferably, the learning may be performed by calculating a fifth value by multiplying the first error by a second weight, multiplying the second error by a third weight, and calculating a sixth value. The error function is generated based on a result obtained by summing six values and the third error.

Advantageously, said categorizing classifies said images based on whether features included in said images are below said threshold.

Preferably, in the classifying step, when it is determined that the features included in the comparison target image are less than or equal to the threshold value, the object included in the comparison target image is classified as the same person as the object included in the reference image. If it is determined that the features included in the threshold value are exceeded, the object included in the comparison target image is classified into an object included in the reference image and another person.

Preferably, prior to the learning step, further comprising the step of editing each image to have a predetermined size based on the point where the first features are expected to be concentrated when the images are input, the learning The step of learning trains the edited images.

In addition, the present invention generates an error function based on the average value of the first features included in the reference image and the average value of the second features located in the class (object) formed in the object unit in the reference image, the error function The present invention proposes an image learning apparatus characterized by learning input images.

The present invention can achieve the following effects through the configuration for achieving the above object.

First, it is possible to classify images of faces through neural network learning.

Second, neural network learning does not need to be repeated every time a new image is input, thereby improving the efficiency of image classification.

1 is a conceptual diagram schematically showing an internal configuration of an image classification system according to an embodiment of the present invention.
2 is a distribution diagram of advanced learned features.
3 is an exemplary view showing the performance of the error function reflecting the global center error proposed in the present invention.
4 is a reference diagram showing the performance of the error functions proposed in the present invention.
5 is a conceptual diagram schematically illustrating an image classification apparatus according to an exemplary embodiment of the present invention.
6 is a flowchart schematically illustrating an image classification method according to a preferred embodiment of the present invention.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. First, in adding reference numerals to the components of each drawing, it should be noted that the same reference numerals are assigned to the same components as much as possible even if displayed on different drawings. In addition, in describing the present invention, when it is determined that the detailed description of the related well-known configuration or function may obscure the gist of the present invention, the detailed description thereof will be omitted. In addition, the following will describe a preferred embodiment of the present invention, but the technical idea of the present invention is not limited thereto and may be variously modified and modified by those skilled in the art.

1 is a conceptual diagram schematically showing an internal configuration of an image classification system according to an embodiment of the present invention.

According to FIG. 1, the image classification system 100 includes an input module 110, an adjustment module 120, a learning module 130, a setting module 140, and a classification module 150.

The input module 110 receives and stores the images to be classified.

The adjustment module 120 edits the images according to a predetermined criterion when the classification target images are input to the input module 110. In one example, the adjustment module 120 may edit each image to have a predetermined size based on the center where the features are expected to be dense in order to efficiently classify each image. In the present invention, the edited image can be used as a mini-batch.

The learning module 130 learns each image edited by the adjustment module 120 and extracts features from each image. The learning module 130 may use a convolutional neural network (CNN) as an example for image learning.

In the present invention, each image is trained using an error function to which global center loss and local center loss are added so that the features are not classified in a specific part so that the images of the face can be classified. Let's do it. This will be described in detail below.

In order to classify similar images such as faces, neural networks need to be distinguished. In the present invention, a global center loss and a local center loss are added to an error function to classify similar images such as faces.

Global center loss and local center loss have been proposed to solve the problem of poor image classification performance because it is difficult to find a manifold that classifies the features extracted from the image. In other words, it is a method of forcing features not to be concentrated at a predetermined point.

Errors used in image classification learning include softmax cross-entropy errors and errors that reduce variance in each class. In the present invention, the combination of these errors and the errors proposed above will be used for image classification learning.

Hereinafter, a description will be given of a face recognition enhancement technique that has undergone discriminative feature learning according to an algorithm in which the errors proposed by the present invention are added to an error function.

Composite product neural networks are a type of deep learning that consists of neural networks. A general neural network processes image data as it is, but a convolutional neural network extracts features from an image and processes the image based on these features.

However, through the ILSVRC (ImageNet Large-Scale Visual Recognition Challenge), we have been able to build a multiplicative neural network that is deeper than before. As a result, the composite product neural network can extract high-dimensional features from an image and improve image recognition performance.

While many deep techniques have been proposed to design deeper neural networks on the premise that deeper convolutional neural networks have better feature representations than shallower convolutional neural networks, the object to be classified has a different nature than the general image. If so, there are additional considerations along with the depth of the neural network.

In the present invention, a model that allows the features of the multiplicative neural network to be further separated from the center of all features while reducing the distance from the center of each class so that similar images such as faces can be classified. Suggest. In the present invention, the method of learning while keeping the distance between classes is defined as a global center loss.

In the present invention, a joint center loss is a concept of an error function including a global center loss and a local center loss.

The model to which the error function based on joint center loss is applied also differs in how images are classified. In case of using the general model, the image is classified by the label derived after learning using softmax, but when using the model to which the error function based on the joint center error is applied, the features extracted from the image and the Compare the distances between the centers and classify the images by selecting the nearest centers of the same class.

The multiplicative neural network learns softmax values by minimizing cross-entropy of the correct label. However, in the face recognition problem, the model trained using only the softmax error has difficulty. This is because features extracted from the model are concentrated together, making it difficult to distinguish between classes.

In order to classify high similarity images such as faces, not only the features to be learned must be classified, but also a model in which the features are distinguished and discriminated, that is, discriminated, is required.

2 is a distribution diagram of deeply learned features. FIG. 2 (a) shows a distribution of distinguishable features, and FIG. 2 (b) shows a distribution of discriminant features.

Learning while forcing the model to have distinctive features yields a high level of recognition in the face recognition field. In addition, models that are forced to have identifying features also have the following advantages in classification:

First, the features extracted from the model are more distinguished than the previous softmax, so the classification is more suitable by measuring the Euclidean distance between the features. In addition, it is possible to find and classify the nearest center by comparing the extracted features with the center of each class.

Second, it is possible to derive features by using multiple images and compare Euclidean distance of each feature to classify them in the order of the most similar images. In particular, it is very practical to extract and classify features without receiving class results. The reason for this is that the model classifies the images by comparing them with the features of the newly added classifications without re-learning each time an image of the new classification is added.

In the present invention, the model is forced to have better discriminating features by using an error function that minimizes the center and distance of each class and an error function that moves the center and distance of all features away. Models that force learning to have better discriminating features can improve the final classification performance by increasing the distance between each feature.

To extract identifying features from the model, there must be an error function that forces the identifying features to be extracted. The present invention proposes an error function based on joint center loss as such an error function. Joint center loss is a concept that includes a global center loss and a local center loss.

First, the global center error will be described, and then the error function based on the joint center error will be described.

The reason for learning to have identifying features is to make sure that the features are not too dense. When the features are dense, the function becomes higher order to satisfy all values, or over fitting occurs, or the function becomes lower order to satisfy all values, resulting in under fitting. This results in a problem of degrading the classification performance of the model.

Training a model through supervised learning is like finding a manifold for each class. However, with dense problems, the larger the number of classes to classify, the more difficult it is to find a suitable manifold compared to a relatively small class. Therefore, it is necessary to easily obtain a suitable manifold by forcing the features not to be extracted in a space that is likely to be dense and difficult to obtain a manifold.

In general, since the space most likely to be concentrated is the center of all the features, in the present invention, as each feature gets closer to the center of all the features, a higher error is given.

The global center error forcing the extracted features not to be concentrated at the center is represented by Equation 1 below.

In the above, c _g means the center of all features, that is, the average of all features, and the present invention defines this as a global center.

The global center may be calculated by averaging coordinate values of all features. The global center may be ideally derived in comparison with all learning data every hour. However, in the present invention, the center of each class is used in a mini-batch for performance reasons.

Learning refers to the process of backpropagation by the error function of neural networks. However, it is practically impossible to use a full batch of all the data you have. Therefore, a method of learning neural networks using some data (ex. 100) and then using some next data (ex. 100) is proposed. Mini-batch refers to a method of learning by using only a part of data in neural network training.

γ means margin. γ is not a variable to be learned as a hyper-parameter, but a value arbitrarily assigned by the user.

x _i means the i th feature extracted from the image. And m is the number of classes. In the present invention, a class is a label utilized for supervised learning, and for example, a dog and a cat are labels if a neural network distinguishing a cat and a dog.

r means the largest distance between the global center and the feature located farthest from the global center. r can be obtained from Equation 2 below.

In the present invention, as shown in Equation 2, the difference between the global center and a specific feature in the image in the mini-batch is found to be the maximum value and is calculated as r.

On the other hand, x _i may mean the entire value derived from the input image or the previous layer.

On the other hand, max (A, B) is a function for deriving a larger value of A and B. In Equation 1, max (0, B) is used so that a value larger than 0 is always derived.

Meanwhile || || means L2 norm.

The global center error L _g is between the first square value (ie, the square of r) and the second square value (ie, the square of the difference between x _i and c _g ) within the mini-batch. It is forced to give an error so that the difference is larger than γ. Minimizing the global center error L _g can force the feature to be larger than γ while learning the feature away from the center. The global center error L _g may thereby force the features extracted from the image not to be centered.

Meanwhile, in the present invention, in order to avoid the inefficiency caused by repeating the learning based on all the data, the global center c _g , the maximum distance r, etc. are calculated using Equation 3, Equation 4, and the like. Update.

Equation 3 shows a gradient of L _g by x _i , and is described in detail as follows.

In Equation 3, the result of ∂L _g / ∂x _i becomes-(x _i -c _g ) or 0 is calculated as max (0, || r || ^2- || x _i -c _g ² + γ).

x _i is far from c _g || x _i -c _g || ² || r || Occurs with a value greater than ² , in which case || r || ^2- || x _i -c _g || ^The phenomenon that ² + γ is smaller than 0 may occur. At this time, there is a problem that the gradient has a negative value and does not normally learn. Equations 1 and 3 are set in such a manner that the result value is not negative in consideration of such a problem.

Equation 4 shows an update equation of c _g , and is described in detail as follows.

3 is an exemplary view showing the performance of the error function reflecting the global center error proposed in the present invention.

In the case of an image of a face, features are generally concentrated in the center, as shown in FIG. When discriminative deep learning is performed by applying a center expansion technique based on an error function reflecting a global center error to such an image, as shown in (b) of FIG. The image can be varied so that the distribution is not dense in the center.

Next, an error function based on joint center loss will be described.

The easiest way to extract identifying features from an image is to learn the variance in each class to be small. If the variance in the class is small, the possibility that the extracted features overlap with other classes is reduced, resulting in the effect that the distinguishing features are derived.

In order to minimize variance in the class, the present invention uses an error function that minimizes the distance between the center of each class and the features of the class to which it belongs. This error function is defined as in Equation 5 below.

In the above, c _yi means the center of the y class, that is, the average of the y class features. Since c _yi is ideally derived compared to all the learning data every hour, but there is no performance efficiency, the present invention finds the center of each class in a mini-batch.

Due to the nature of neural network learning, we update it with variables to satisfy the entire data set. c _yi and c _g are initially initialized to a random value or 0 and then continuously updated to a suitable value (a value that satisfies all the data).

Since L _g has been described above, its detailed description will be omitted here.

As in the case of the global center error, the joint center error also uses the global center c _g , Equation 7, etc. to avoid the inefficiency caused by repeating the learning based on all the data. Update the largest distance r and so on.

Equation 6 shows a gradient of L _c by x _i , and is described in detail as follows.

Equation 7 shows an update equation of c _j , and is described in detail as follows.

In the above, δ (y _i = j) is a function that can derive only a value of 0 or 1 as a Direc function. That is, δ is 1 when the internal condition is satisfied and 0 when the internal condition is not satisfied. In the present invention, δ derives 1 when the operation in () is true and 0 when the operation in () is false. This is to ensure that the center of your class can only affect the characteristics of your class.

c _j stands for the center of each class. Δc _j modifies c _j corresponding to each class by using the difference between the features extracted from the image and the center of the class to which the features belong. Utilizing the fact △ c _j and when updating the c _j balances utilize a hyper parameter α.

The hyper parameter α is arbitrarily set so that the center of the class is not easily changed by the training data, and is a value for controlling an imbalance caused during learning. In the present invention, α may have a value greater than 0 and may be changed by an optimizer used for learning.

However, you cannot train the model using only L _c and L _g .

The final second error function is the sum of the L _c , L _g , and softmax cross-entropy errors, which are described in Equation 8 below.

In the above description, W and b mean a factor to be learned. T at the top right of W means transpose for matrix multiplication, and j or yi at the bottom right describes the number of learning factors. In particular, j means the index of the column of the feature, and yi means the class of the i th.

The equation corresponding to L _s is generally used to learn neural networks and is called softmax loss, cross-entropy, or negative likelihood. e means exponential.

In the present invention, when calculating an error function based on the joint center error, a gradient descent technique used in a normal learning process may be used.

The error due to L _c decreases exponentially as it approaches c _yi , and the error due to L _g decreases exponentially as it moves away from c _g . Therefore, it is possible to mitigate that the feature of the model is too close to the center of each feature or the feature is abnormally far from the center of the feature. In other words, using Equation 8, as the feature to be learned becomes far from the center of each class, the error is increased to become an identifying feature, and the feature around c _g , which is a part that is likely to generate a contact point, increases the error to other classes. You can force learning to avoid crowding the liver.

λ _c and λ _g are not variables that can be learned with hyper-parameters, but values that the user arbitrarily specifies. In the present invention, λ _c and λ _g may have a value greater than zero and are used to adjust the influence of two error functions. That is, as λ _c and λ _g increase, the feature distribution gets closer to the center of its class. If it is 0, the result is the same as the model trained using only softmax.

In the present invention, as described in Equation 8, by adding two, such as c _g and the difference between the features, c _yi and the difference between the features as an error function, each feature reduces the variance in its class and at the same time Study away from the center.

N is the number of classes.

4 is a reference diagram showing the performance of the error functions proposed in the present invention. In (a) to (d) of FIG. 4, each point means a feature derived from the learned neural network, and points of different colors mean features belonging to different classes.

4 (a) shows a distribution of features extracted as a result of learning using an error function that is considered a global center loss, that is, a first error function. In this case, the first error function is an error function in which a softmax cross-entropy error and a global center loss are reflected.

As shown in (a) of FIG. 4, when using the first error function proposed in the present invention, the features extracted from the images can be distributed so as not to be concentrated at the center, thereby making it possible to classify the images of the faces.

4 (b) to (d) show distributions of features extracted as a result of learning using an error function, that is, a second error function in which joint center loss is considered. In this case, the second error function is an error function in which the softmax cross entropy error, the global center error, and the joint center loss are reflected.

In detail, FIG. 4 (b) shows the distribution of features obtained by setting the ratio of λ _c and λ _g to 1: 1, and FIG. 4 (c) shows the ratio of λ _c and λ _g to 0.1: 1. The distribution of the characteristics obtained by setting is shown. 4D shows the distribution of features obtained by setting the ratio of λ _c and λ _g to 0.01: 1.

In the case of using the second error function proposed in the present invention, the features extracted from the image can be distributed so as not to be concentrated at the center, as in the case of using the first error function, thereby making it possible to classify the image of the face.

This will be described with reference to FIG. 1 again.

After the neural network training by the learning module 130 is finished, the setting module 140 performs a function of setting a threshold value based on the features extracted from each image.

For example, the setting module 140 may set a threshold in the following order.

First, the setting module 140 derives the features by using a test dataset in the learned neural network.

The setting module 140 then temporarily sets an arbitrary value as a threshold.

The setting module 140 then measures face recognition accuracy. This function of the configuration module 140 can be repeated several times.

The setting module 140 then determines the most appropriate threshold value through repeated measurements of face recognition.

The classification module 150 performs a function of classifying each image based on the threshold set by the setting module 140. In one example, the classification module 150 may classify each image in the following order.

First, the classification module 150 determines whether a feature of an image is equal to or less than a threshold. This function of the classification module 150 is represented by a formula as follows.

D (x _i , x _j ) ≤ d

In the above, D (x _i , x _j ) means a feature extracted from the image. And D means a threshold value.

If it is determined that the feature of the image is less than or equal to the threshold, the classification module 150 then classifies the image as being the same as the image associated with the threshold. On the other hand, if it is determined that the feature of the image is above the threshold, the classification module 150 then classifies the image as not being the same person as the image associated with the threshold (ie, being another person).

As mentioned above, one Embodiment of this invention was described with reference to FIGS. Hereinafter, the preferable form of this invention which can be inferred from such one Embodiment is demonstrated.

5 is a conceptual diagram schematically illustrating an image classification apparatus according to an exemplary embodiment of the present invention.

According to FIG. 5, the image classifying apparatus 200 includes an image learner 210, a threshold value setting unit 220, an image classifying unit 230, a power supply unit 240, and a main control unit 250.

The power supply unit 240 performs a function of supplying power to each component of the image classification apparatus 200. The main controller 250 performs a function of controlling the overall operation of each component of the image classification apparatus 200. The image classification apparatus 200 may be provided in an apparatus for processing images to classify the images. Since the power supply unit and the main control unit also exist in the apparatus for processing the image, in view of this point, the power supply unit 240 and the main control unit 250 may not be provided in this embodiment.

The image learner 210 calculates an error function based on the information about the first features included in the reference image and the information about the second features located in a class formed in units of objects in the reference image. Perform the function to create. In addition, the image learner 210 performs a function of learning the input images based on the generated error function. The image learner 210 is a concept corresponding to the learning module 130 of FIG. 1.

The image learner 210 may use a convolutional neural network when learning the images.

The image learner 210 based on a first error generated based on the average value of the first features, a second error generated based on the average value of the second features, and a third error related to cross-entropy. An error function can be generated. In detail, the image learner 210 calculates a fifth value by multiplying the first error by a second weight, multiplies the second error by a third weight, and calculates a sixth value, and the fifth value and the sixth value. And an error function based on the result obtained by summing the third errors. In the above description, the first error is a concept corresponding to L _g , the second error is a concept corresponding to L _c , and the third error is a concept corresponding to L _s .

When calculating the first error, the image learner 210 may calculate the first error based on first difference values between the average value of the first features and each feature included in the reference image, and the first weight. have. In detail, the image learner 210 applies a maximum value selected from the first difference values to L2-norm to calculate a first value, and applies each first difference value to L2-norm to a second value. The first error may be calculated based on the results obtained by comparing the third values obtained by adding the first weight to the first value and subtracting each second value and 0. FIG.

When calculating the second error, the image learner 210 based on the second difference between the average value of the second features and each feature included in the reference image, and the number of classes included in the reference image, is determined. The error can be calculated. In detail, the image learning unit 210 calculates second difference values for each class, calculates fourth values for each class by applying the second difference values calculated for each class to L2-norm, and for each class. The second error may be calculated based on the result obtained by summing the calculated fourth values.

The threshold setting unit 220 performs a function of setting a threshold value based on a result obtained by learning the images through the image learning unit 210. The threshold setting unit 220 is a concept corresponding to the setting module 140 of FIG. 1.

The image classifying unit 230 classifies the images based on the threshold set by the threshold setting unit 220. The image classifier 230 is a concept corresponding to the classification module 150 of FIG. 1.

The image classifier 230 may classify the images based on whether the features included in the images are below a threshold. In detail, if it is determined that the features included in the comparison target image are less than or equal to the threshold value, the image classification unit 230 classifies the objects included in the comparison target image into the same person as the objects included in the reference image. On the other hand, if it is determined that the features included in the comparison target image exceeds the threshold, the image classifying unit 230 classifies the objects included in the comparison target image into objects and others included in the reference image.

The image classification apparatus 200 may further include an image adjusting unit 260.

The image adjusting unit 260 edits each image to have a predetermined size based on a point where the first features are expected to be concentrated when the images are input. In this case, the image learner 210 may learn the images edited by the image adjuster 260. The image adjusting unit 260 is a concept corresponding to the adjusting module 120 of FIG. 1.

Meanwhile, in the present exemplary embodiment, the image learner 210 may be configured independently from the image classifier 200 and implemented as an image learner.

Next, a method of operating the image classification apparatus 200 will be described.

6 is a flowchart schematically illustrating an image classification method according to a preferred embodiment of the present invention.

First, the image learner 210 based on the information about the first features included in the reference image and the information about the second features located in a class formed in units of objects in the reference image. To generate and learn the input images based on the error function (S310).

Thereafter, the threshold setting unit 220 sets a threshold based on the result obtained by learning the images (S320).

The image classifier 230 classifies the images based on the threshold value (S330).

Meanwhile, before operation S310, the image adjusting unit 260 may edit each image to have a predetermined size based on a point where the first features are expected to be concentrated when the images are input.

Although all components constituting the embodiments of the present invention described above are described as being combined or operating in combination, the present invention is not necessarily limited to these embodiments. In other words, within the scope of the present invention, all of the components may be selectively operated in combination with one or more. In addition, although all of the components may be implemented in one independent hardware, each or some of the components of the components are selectively combined to perform some or all of the functions combined in one or a plurality of hardware It may be implemented as a computer program having a. In addition, such a computer program is stored in a computer readable medium such as a USB memory, a CD disk, a flash memory, and the like, and is read and executed by a computer, thereby implementing embodiments of the present invention. The recording medium of the computer program may include a magnetic recording medium, an optical recording medium and the like.

In addition, all terms including technical or scientific terms have the same meaning as commonly understood by a person of ordinary skill in the art unless otherwise defined in the detailed description. Terms used generally, such as terms defined in a dictionary, should be interpreted to coincide with the contextual meaning of the related art, and shall not be interpreted in an ideal or excessively formal sense unless explicitly defined in the present invention.

The above description is merely illustrative of the technical idea of the present invention, and various modifications, changes, and substitutions may be made by those skilled in the art without departing from the essential characteristics of the present invention. will be. Accordingly, the embodiments disclosed in the present invention and the accompanying drawings are not intended to limit the technical spirit of the present invention but to describe the present invention, and the scope of the technical idea of the present invention is not limited by the embodiments and the accompanying drawings. . The scope of protection of the present invention should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the present invention.

Claims (18)

Generate an error function based on the information about the first features included in the reference image and the information about the second features located in the corresponding class among the classes formed in units of objects in the reference image. An image learning unit learning the input images based on the error function;
A threshold setting unit for setting a threshold value based on a result obtained by learning the input images; And
And an image classifier classifying the input images based on the threshold value.
The image learner may include a first error calculated based on an average value of the first features, a second error calculated based on an average value of the second features, and a cross entropy to increase the distance between the classes. calculate the error function based on a third error associated with
The image learning unit may include a predetermined number of classes to be classified; And a distance from an average value of the first features to a feature that is farthest from the average value of the first features among the first features included in the reference image. And learning models associated with the features included in the reference image by using the error function to learn the input images. The method of claim 1,
And the image learner learns the input images using a convolutional neural network. delete The method of claim 1,
And the image learner calculates the first error based on first difference values between the average value of the first features and each feature included in the reference image, and a first weight. The method of claim 4, wherein
The image learner calculates a first value by applying a maximum value selected from the first difference values to Euclidean norm, and calculates second values by applying each first difference value to Euclidean norm, And calculating the first error based on the results obtained by comparing the first values with zeros and the third values obtained by subtracting each of the second values and zero. The method of claim 1,
The image learner calculates the second error based on second difference values between the average value of the second features and each feature included in the reference image, and the number of classes included in the reference image. Image classification apparatus. The method of claim 6,
The image learning unit calculates the second difference values for each class, calculates fourth values for each class by applying the second difference values calculated for each class to Euclid norm, and adds the fourth values calculated for each class. And calculating the second error based on the result obtained. The method of claim 1,
The image learner calculates a fifth value by multiplying the first error by a second weight, and calculates a sixth value by multiplying the second error by a third weight, and the fifth value, the sixth value, and the fifth value. And an error function is generated based on a result obtained by summing 3 errors. The method of claim 1,
And the image classifier classifies the input images based on whether the features included in the input images are less than or equal to the threshold value. The method of claim 9,
If it is determined that the features included in the comparison target image are less than or equal to the threshold value, the image classification unit classifies the object included in the comparison target image into the same person as the object included in the reference image, and the features included in the comparison target image And determining that the object included in the comparison target image is classified into an object included in the reference image and another person when it is determined that the threshold value is exceeded. The method of claim 1,
And an image adjusting unit for editing each input image to have a predetermined size based on a point where the first features are expected to be concentrated when the images are input.
And the image learning unit learns the edited input images. Generate an error function based on the information about the first features included in the reference image and the information about the second features located in the corresponding class among the classes formed in units of objects in the reference image. Learning the input images based on the error function;
Setting a threshold value based on a result obtained by learning the input images; And
Classifying the input images based on the threshold value;
The training of the input images may include: a first error calculated based on the average value of the first features, a second error calculated based on the average value of the second features, and so as to distance the classes; Calculating the error function based on a third error related to cross-entropy, wherein a predetermined number of classes to be classified; And a distance from an average value of the first features to a feature that is farthest from the average value of the first features among the first features included in the reference image. And learning the input images using an error function so that a learning model associated with features included in the reference image has an identifying feature. delete The method of claim 12,
The training may include calculating the first error based on first difference values between a mean value of the first features and each feature included in the reference image, and a first weight. The method of claim 12,
The learning may include calculating the second error based on the second difference between the average value of the second features and each feature included in the reference image, and the number of classes included in the reference image. An image classification method characterized by the above-mentioned. The method of claim 12,
The classifying step may include classifying the input images based on whether features included in the input images are less than or equal to the threshold value. The method of claim 12,
Editing each input image to have a predetermined size based on a point where the first features are expected to be concentrated when the images are input;
The training may include learning the edited input images. Generate an error function based on an average value of first features included in a reference image and an average value of second features located in a corresponding class among classes formed in units of objects in the reference image; Learn the input images based on the function,
A first error calculated based on an average value of the first features, a second error calculated based on an average value of the second features, and a cross-entropy related to cross-entropy so as to distance the classes between the classes; Calculating the error function based on the error, and classifying a predetermined number of classes; And a distance from an average value of the first features to a feature that is farthest from the average value of the first features among the first features included in the reference image. And learning the input images by using an error function so that a learning model associated with features included in the reference image has an identifying feature.

Images