JPWO2020065908A1 - Pattern recognition device and pattern recognition method - Google Patents

Abstract

The pattern recognition device 20 is the first learning including the feature amount data output by one of the plurality of layers of the neural network type classifier in which the plurality of layers into which the image data for training is input are connected in a layered manner. The data and the second training data including the feature amount data output by one layer different from one layer of the plurality of layers of the neural network type classifier in which the image data for learning different from the image data is input are A prediction unit 21 that predicts the discrimination performance of the neural network type classifier after being used and learned, and a determination unit 22 that determines the range of layers to be learned of the neural network type classifier based on the predicted discrimination performance. And.

Description

The present invention relates to a pattern recognition device, a pattern recognition method and a pattern recognition program, and more particularly to a pattern recognition device, a pattern recognition method and a pattern recognition program to which a statistical pattern recognition technique is applied.

Image recognition technology using deep learning is installed in various video surveillance systems. Deep learning deals with calculation algorithms such as neural networks.

Patent Documents 1 to 4 describe techniques related to neural networks. For example, Patent Document 1 describes an information processing apparatus using a multi-layer neural network. Further, Patent Document 2 describes a machine learning device capable of creating a neural network of an appropriate scale.

Further, Patent Document 3 describes a neural network learning device capable of optimizing the structure of a neural network. Further, Patent Document 4 describes a character recognition device using a neuro that learns not only the overall shape of a character but also the shape of each small area of the character and recognizes the character.

As a procedure for executing deep learning, a procedure called fine-tuning is often used. In fine tuning, a pre-trained dictionary learned by using a large amount of image data in advance is used as an initial value, and image data indicating an object originally recognized by the video surveillance system is additionally learned. ..

The dictionary contains network models, such as neural networks, and weight parameters for network models. The weight parameter of the network model is modified by additionally learning the image data indicating the object.

For example, Non-Patent Document 1 describes a technique related to fine tuning. Non-Patent Document 1 describes a learning method in which a feature extraction network is constructed by unsupervised learning and then the weight parameters of the entire network are readjusted by fine tuning using a recognition target label.

Further, Non-Patent Document 2 describes a method of modifying the network weight parameter when new training data is obtained, and a network weight accompanied by addition of neurons to the output layer when new class data is added. Describes how to modify the parameters.

When the training data or the class to be recognized is added, there is a method of modifying the weight parameter of the entire network and a method of modifying only the weight parameter of the middle layer of the 2nd to 3rd layers close to the output layer of the network.

The method of modifying the weight parameter of the entire network has the advantage that it is theoretically more likely to construct a network with a high discrimination rate than the method of modifying only the weight parameter of the middle layer close to the output layer of the network. Have.

Further, the method of modifying only the weight parameter of the intermediate layer close to the output layer of the network has a feature that the amount of calculation related to the modification of the weight parameter is smaller than the method of modifying the weight parameter of the entire network.

The reason why it is more likely that a network with a high discrimination rate can be constructed by modifying the weight parameter of the entire network is that the appropriate extraction of features from newly added learning data and the class to be recognized is the first stage of the network. This is because it is more likely to be executed in the part.

Therefore, when both the computational resources and the training data are sufficiently available, it is preferable to modify the weight parameters of the entire network when the training data and the class to be recognized are added. Also, if at least one of the computational resources and the training data is not sufficiently available, it is preferable to modify only the weight parameter of the layer close to the output layer of the network when the training data or the class to be recognized is added.

In addition, the learning data is often image data. However, in the learning of SVM (Support Vector Machine) and LVQ (Learning Vector Quantization) classifiers, the feature extraction means are completely independent of the discriminating means. Therefore, a single feature amount data output by the feature extraction means is often used as learning data of the identification means.

The feature amount data has an advantage that the privacy problem is less likely to occur even if it is used as compared with the image data. That is, it is conceivable that the learning device is required to be able to record the feature amount data abstracted to the extent that the individual is not specified, although the image data cannot be recorded.

When the above requirement is imposed on the learning device, it is conceivable that there may be a situation in which only the feature amount data exists in the learning device and the image data corresponding to the feature amount data does not exist because it is not recorded. In the above situation, learning data in which each feature amount data extracted from different image data in different layers of the neural network type classifier may be mixed may be used.

Japanese Unexamined Patent Publication No. 2018-026040 Japanese Unexamined Patent Publication No. 2017-182319 Japanese Unexamined Patent Publication No. 2017-037392 Japanese Unexamined Patent Publication No. 07-160830

P. Sermanet, K. Kavukcuoglu, S. Chintala, and Y. LeCun, "Pedestrian detection with unsupervised multi-stage feature learning," In Proc. International Conference on Computer Vision and Pattern Recognition (CVPR), 2013. Christoph K¨ading, Erik Rodner, Alexander Freytag, and Joachim Denzler, "Fine-tuning Deep Neural Networks in Continuous Learning Scenarios," In ACCV 2016 Workshop on Interpretation and Visualization of Deep Neural Nets, 2016.

The training method in which the training data in which the image data and the feature amount data are mixed or the training data in which the feature amount data extracted from different image data in different layers of the neural network type classifier is mixed is used is used. There are challenges. Hereinafter, the reason why there is a problem in the learning method will be described with reference to the drawings.

FIG. 11 is an explanatory diagram showing an example of a neural network type classifier. FIG. 11 shows the structure of the neural network type classifier and the position of the layer from which the feature data is extracted. The neural network type discriminator shown in FIG. 11 is a network type discriminator in which a convolution layer and a fully connected layer are connected in a layered manner.

In the convolution layer shown in FIG. 11, convolution is performed in a local region. Further, each number in the convolution layer shown in FIG. 11 indicates a size obtained by multiplying a vertical size and a horizontal size of data represented in two dimensions such as an image or a feature amount. For example, "224x224" is the size of data with a vertical size of 224 (pixels or pieces) and a horizontal size of 224 (pixels or pieces).

In addition, the neurons in the fully connected layer shown in FIG. 11 are connected to all the neurons in the previous layer. Further, each number in the fully connected layer shown in FIG. 11 indicates the size of the feature amount data expressed in one dimension. For example, "4096" means that the size of the feature data output from the layer is 4096 (pieces).

When the network learning shown in FIG. 11 is executed, particularly when fine tuning is executed, it is considered that a large amount of training data that can be used is preferable. Also, from the viewpoint of adapting the network to new data as much as possible, it is considered preferable to modify the weight parameter of the network as much as possible by fine tuning.

However, when the network weight parameter is modified, the feature data extracted from the layer above the layer having the modified weight parameter becomes unavailable. Therefore, when the learning data including the feature amount data is used, there is a problem that the available learning data is reduced by the amount that the weight parameter is changed.

As described above, increasing the number of layers in which the weight parameter is modified by fine tuning and increasing the amount of reusable feature data are not satisfied at the same time. Therefore, there is a need for a method of optimally determining the layer in which the weight parameter is modified while considering the amount of reusable feature data. Patent Document 1 to Patent Document 4 and Non-Patent Document 1 to Non-Patent Document 2 do not describe the above-mentioned determination method.

Therefore, an object of the present invention is to provide a pattern recognition device, a pattern recognition method, and a pattern recognition program capable of performing fine tuning while considering the amount of reusable feature amount data, which solves the above-mentioned problems. ..

The pattern recognition device according to the present invention includes feature data output by one of a plurality of layers of a neural network type classifier in which a plurality of layers into which image data for learning is input are connected in a layered manner. Second training data including one training data and feature amount data output by one layer different from one of the plurality of layers of the neural network type classifier in which image data for learning different from the image data is input. A predictive means for predicting the discrimination performance of the neural network type classifier after learning using and, and a determination means for determining the range of the layer to be learned of the neural network type classifier based on the predicted discriminant performance. It is characterized by having and.

The pattern recognition device according to the present invention is a neural network type discriminator in which a plurality of layers are connected in a layered manner. The learning means for learning the neural network type classifier using the training data including the feature amount data output by the candidate layer in the determined range of the classifier, and the discriminating performance of the neural network type classifier after learning. The evaluation means to be evaluated, the storage means to store the parameters of the trained and derived neural network type classifier together with the candidates of the determined range, the evaluated discrimination performance and the number of training data used for training. It is characterized by including a selection means for selecting a parameter from a storage means based on the basis.

The pattern recognition method according to the present invention includes feature data output by one of a plurality of layers of a neural network type classifier in which a plurality of layers into which image data for learning is input are connected in a layered manner. Second training data including one training data and feature amount data output by one layer different from one of the plurality of layers of the neural network type classifier in which image data for learning different from the image data is input. It is characterized in that the discrimination performance of the neural network type classifier after learning is predicted by using and, and the range of the layer to be learned of the neural network type classifier is determined based on the predicted discrimination performance. ..

In the pattern recognition method according to the present invention, a candidate for a range of layers to be learned of a neural network type classifier in which a plurality of layers are connected in layers is determined, and an image data for learning is input to the neural network type classifier. The neural network type classifier is trained using the training data including the feature amount data output by the candidate layer in the determined range, and the discriminating performance of the neural network type classifier after the training is evaluated and trained. The derived neural network type classifier parameters are stored in the storage means together with the candidates in the determined range, and the parameters are selected from the storage means based on the evaluated discrimination performance and the number of training data used for training. It is characterized by that.

The pattern recognition program according to the present invention is feature data output by one of a plurality of layers of a neural network type classifier in which a plurality of layers in which image data for learning is input are connected in a layered manner. The first training data including 2 Prediction processing that predicts the discrimination performance of the neural network type classifier after learning using the training data, and the range of the layer to be trained of the neural network type classifier is determined based on the predicted discrimination performance. It is characterized in that the decision processing to be performed is executed.

The pattern recognition program according to the present invention is a neural network type discriminator in which a plurality of layers are connected in a layered manner. Learning process to learn the neural network type classifier using the training data including the feature amount data output by the candidate layer in the determined range of the network type classifier, and the discrimination performance of the neural network type classifier after learning. Evaluation processing that evaluates the parameters of the neural network type classifier that has been trained and derived, and storage processing that stores the parameters of the neural network type classifier in the storage means together with the candidates in the determined range, and the evaluated discrimination performance and the training data used for training. It is characterized in that a selection process of selecting a parameter from a storage means based on a number is executed.

According to the present invention, fine tuning can be performed while considering the amount of reusable feature data.

It is a block diagram which shows the structural example of the 1st Embodiment of the pattern recognition apparatus by this invention. It is a flowchart which shows the operation of the fine tuning execution processing by the pattern recognition apparatus 100 of 1st Embodiment. It is a block diagram which shows the structural example of the 2nd Embodiment of the pattern recognition apparatus by this invention. It is explanatory drawing which shows the example of the learning data stored in the learning data storage means 202. It is explanatory drawing which shows the example of the relationship between the range of the layer to which fine tuning is executed, and the amount of training data. It is a flowchart which shows the operation of the fine tuning execution processing by the pattern recognition apparatus 200 of 2nd Embodiment. It is explanatory drawing which shows the other example of the learning data stored in the learning data storage means 202. It is explanatory drawing which shows the hardware configuration example of the pattern recognition apparatus by this invention. It is a block diagram which shows the outline of the pattern recognition apparatus by this invention. It is a block diagram which shows the other outline of the pattern recognition apparatus by this invention. It is explanatory drawing which shows an example of the neural network type classifier.

[Description of configuration]
Embodiment 1.
Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing a configuration example of a first embodiment of the pattern recognition device according to the present invention.

As shown in FIG. 1, the pattern recognition device 100 includes a neural network type classifier 101, a first learning data storage means 102, a second learning data storage means 103, and a learning means 104.

As described above, when there are a plurality of feature data in which each output from a plurality of intermediate layers of the neural network type classifier is recorded in addition to the image data as the training data for fine tuning, the discrimination performance and the feature There is a trade-off relationship with the reusability of quantitative data.

The discrimination performance of the neural network type classifier and the reusability of the feature data are related to the range of the layer to be fine-tuned. In the pattern recognition device 100 of the present embodiment, when image data and feature data, or a plurality of types of feature data are given as training data, both the discrimination performance after fine tuning and the reusability of the training data are available. Determine the range of layers to be fine-tuned in consideration of.

The neural network type classifier 101 is a classifier that performs pattern recognition processing using a neural network.

A neural network is composed of, for example, a convolution layer in which convolution is performed in a local region, a pooling layer in which a value of a specified property is extracted in the local region, and neurons connected to all neurons in the previous layer. It is a network in which fully connected layers and the like are connected in layers.

The neural network type classifier 101 of the present embodiment is, for example, the neural network type classifier shown in FIG. Further, the neural network type classifier 101 may be ResNet, GoogleNet, MobileNet, or the like.

The first learning data storage means 102 has a function of storing the first learning data. The first training data corresponds to, for example, the feature amount data output from one layer when the training image data is input to the neural network type classifier 101, and the input training image data. The correct answer class and the identification information of the layer from which the feature data is output are included.

The first training data of the present embodiment is generated in advance by the neural network type classifier 101. For example, when image data for learning is input to the input layer of the neural network type classifier shown in FIG. 11, the values of neurons in the effective state (firing) of _{the intermediate layer L 3 are sampled to generate feature data.} Will be done. The neural network type classifier 101 generates the first training data based on the generated feature data.

Then, the first learning data storage unit 102, first comprises the feature amount data generated, and information indicating the correct class corresponding to the image data for the inputted learning, and information indicating the intermediate layer L ₃ 1 The training data is saved. If the predetermined layer is an input layer, saving the feature amount data is equivalent to saving the input image data for learning.

The second learning data storage means 103 has a function of storing the second learning data. The second training data includes, for example, features that are output when image data different from the training image data that is the source of the feature data included in the first training data is input to the neural network type classifier 101. Includes quantity data. The feature amount data included is data output from one layer different from the one layer in which the feature amount data included in the first learning data is output.

Further, the second learning data includes the correct answer class corresponding to the input image data and the identification information of the layer from which the feature amount data is output together with the feature amount data.

The second learning data is data having the same structure as the first learning data. As described above, the second training data includes the feature data sampled from the layer in which the feature data included in the first training data is sampled and the feature data sampled from another layer.

Further, the image data for learning used for generating the first learning data stored in the first learning data storage means 102 is used for generating the second learning data stored in the second learning data storage means 103. It is different from the image data for learning used. The first learning data and the second learning data generated based on the common image data for learning may be stored in each storage means.

The learning means 104 has a function of performing fine tuning on the neural network type classifier 101 by using both the first learning data and the second learning data. Hereinafter, some methods of executing fine tuning by the learning means 104 will be shown.

For example, the feature amount data F ₁ from the intermediate layer L ₁ shown in FIG. 11, the feature data F ₂ from the intermediate layer L ₂ is extracted, respectively. The feature data F ₁ is included in the first training data, and the feature data F ₂ is included in the second training data. The learning means 104 may perform fine tuning only on the layer higher than _{the intermediate layer L 1} by using both the first learning data and the second learning data.

When performing fine tuning only on the layers above the intermediate layer L _{1 , the learning means 104 inputs the feature data F 1} into the intermediate layer L ₁ and the feature data F ₂ into the intermediate layer L ₂ . do. The learning means 104 reduces the value of the error function by back propagation, for example, using the sum of squares of the output from the output layer and the correct answer label as the error function, as in the case where the image data is input to the input layer. Modify the network weighting parameter to.

Learning means 104 stops the reverse propagation of error in backpropagation an intermediate layer L _1. According to the above method, the layer in which the weight parameter is corrected by fine tuning is limited to the layer higher than _{the intermediate layer L 1.} However, the feature data F ₁ and the feature data F ₂ become data that can be used in the neural network type classifier 101 after the fine tuning is executed.

Further, the learning means 104 learns _{only the layer higher than the intermediate layer L 1} by the above method, and then learns the layer higher than the intermediate layer L _{2 using the feature data F 2.} Fine tuning may be performed.

According to the method of performing fine tuning in two steps, the learning means 104 can adapt the learning data up _{to the lower intermediate layer L 2 of the neural network type classifier 101.}

However, once the weight parameters of the layers above the _{intermediate layer L 2} are modified, the feature data (feature data F _{1 in this example) sampled from the layers above the intermediate layer L 2} becomes a neural network type. It becomes unusable in the classifier 101. If there is a request to avoid the reduction of training data, this method does not meet the request.

When performing fine tuning on the neural network type classifier 101 using the first learning data and the second learning data, the learning means 104 is the target of fine tuning after conducting the above examination. Determine the range of layers.

In the above example, the learning means 104 learns only the layer higher than the intermediate layer L _{1 by using the feature data F 1} and the feature data F ₂ , or further learns only the feature data F ₂ . Whether or not to train up to the intermediate layer L _{2 by} using is determined in consideration of each discrimination performance and the amount of decrease in training data.

The above-mentioned discrimination performance is a value that can be predicted before the learning means 104 performs fine tuning. After determining the range of the layer to be fine-tuned for the neural network type classifier 101, the learning means 104 executes fine-tuning for the determined range of the neural network type classifier 101.

For example, the learning means 104 may determine the range in which the maximum score is calculated among the scores calculated by the following formulas in the range of the layer to be executed for fine tuning.

(Score) = (Discrimination performance) -α × (Reduction amount of learning data) ... Equation (1)

However, the constant α in the equation (1) is a constant for aligning the dimension of the discrimination performance with the dimension of the reduction amount of the learning data.

Further, the learning means 104 performs learning by using the first learning data and the second learning data. For example, the first training data corresponds to the feature amount data output from a predetermined layer when the training image data is input to the neural network type classifier 101 and the input training image data. Includes a set of the correct answer class to be used and the identification information of the layer from which the feature data is output.

The feature amount data included in the first training data and the feature amount data included in the second learning data of the present embodiment are not basically data generated based on the same image data. Therefore, the learning means 104 of the present embodiment is different from a general learning means that generates a plurality of types of feature data for the same image data, combines the generated feature data, and then learns the classifier. ..

Further, the learning means 104 of the present embodiment uses feature amount data extracted from two different layers. However, the learning means 104 may also use the feature amount data extracted from each of more layers for learning.

Normally, the teacher information is given as an ideal output value from the output layer, which is the uppermost layer. The learning means 104 updates the weight parameter of the network between layers from the upper layer side so as to minimize the error between the output value from the output layer of the neural network type classifier 101 and the ideal output value.

That is, the learning means 104 executes fine tuning in a procedure in which the residuals are propagated layer by layer to the lower layers. Therefore, the learning means 104 of the present embodiment does not perform fine tuning only on the intermediate layer.

[Explanation of operation]
Hereinafter, an operation of executing fine tuning of the pattern recognition device 100 of the present embodiment will be described with reference to FIG. FIG. 2 is a flowchart showing the operation of the fine tuning execution process by the pattern recognition device 100 of the first embodiment.

First, the image data for learning is input to the neural network type classifier 101. Next, feature data is output from one predetermined layer of the neural network type classifier 101.

Next, the neural network type classifier 101 outputs the output feature amount data, the information indicating the correct answer category corresponding to the input learning image data, and the feature amount data such as the number of layers. The first training data including the information indicating the position of the layer is generated (step S101). The neural network type classifier 101 stores the generated first learning data in the first learning data storage means 102.

Next, learning image data different from the learning image data input in step S101 is input to the neural network type classifier 101. Next, the feature amount data is output from a layer different from the predetermined layer for the first training data of the neural network type classifier 101.

Next, the neural network type classifier 101 outputs the output feature amount data, the information indicating the correct answer category corresponding to the input learning image data, and the feature amount data such as the number of layers. A second training data including information indicating the position of the layer is generated (step S102). The neural network type classifier 101 stores the generated second learning data in the second learning data storage means 103.

Next, the learning means 104 determines the range of the layer to be fine-tuned for the neural network type classifier 101 (step S103). The learning means 104 is determined by using both the first learning data stored in the first learning data storage means 102 and the second learning data stored in the second learning data storage means 103.

Specifically, the learning means 104 uses the first learning data and the second learning data to learn only the layer higher than the predetermined layer, or further uses only the second learning data to learn the lower layer. Whether to learn up to the layer of is determined in consideration of the discrimination performance after each learning and the amount of decrease in the learning data.

For example, the learning means 104 predicts the discrimination performance after learning, and calculates the score of the equation (1) based on the predicted discrimination performance. The learning means 104 determines the range of the layer to be executed based on the calculated score of the formula (1).

After the determination, the learning means 104 performs fine tuning on the determined range of the neural network type classifier 101 (step S104). After executing the fine tuning, the pattern recognition device 100 ends the fine tuning execution process.

[Explanation of effect]
The pattern recognition device 100 of the present embodiment includes a neural network type classifier 101 represented by a network structure. Further, the pattern recognition device 100 includes a first learning data storage means 102 for storing the first learning data and a second learning data storage means 103 for storing the second learning data.

The first training data includes feature amount data output from one layer when the training image data is input to the neural network type classifier 101, and a correct answer class corresponding to the input training image data. , Includes the identification information of the layer that outputs the feature amount data.

The second training data is the feature amount data included in the first training data when the image data different from the training image data used for generating the first training data is input to the neural network type classifier 101. Includes feature amount data output from one layer that is output and a layer that is different from the output layer. Further, the second learning data includes a correct answer class corresponding to the input image data for learning and identification information of the layer that outputs the feature amount data.

Further, the pattern recognition device 100 includes a learning means 104 that executes fine tuning on the neural network type classifier 101 by using both the first learning data and the second learning data.

The pattern recognition device 100 of the present embodiment can determine the optimum range of the layer in which the weight parameter is corrected by fine tuning. The reason is that the learning means 104 calculates an index representing the result of fine tuning from the two viewpoints of discrimination performance after fine tuning and reusability of learning data, and determines the optimum learning range based on the calculated index. To choose.

Embodiment 2.
[Description of configuration]
Next, a second embodiment of the pattern recognition device according to the present invention will be described with reference to the drawings. FIG. 3 is a block diagram showing a configuration example of a second embodiment of the pattern recognition device according to the present invention.

As shown in FIG. 3, the pattern recognition device 200 includes a neural network type classifier 201, a learning data storage means 202, a learning means 203, a learning data selection means 204, a relearning range determination means 205, and an evaluation means. It includes 206, a re-learning result storage means 207, and a re-learning result selection means 208.

The neural network type classifier 201 is a classifier that performs pattern recognition processing using a neural network. The function of the neural network type classifier 201 is the same as the function of the neural network type classifier 101 of the first embodiment.

The learning data storage means 202 has a function of storing learning data including image data for learning and feature amount data output from the intermediate layer of the neural network type classifier 201. The stored learning data includes the first learning data of the first embodiment and the second learning data. The training data may include only one of the image data and the feature amount data.

The learning means 203 has a function of executing fine tuning on the neural network type classifier 201 by using the learning data stored in the learning data storage means 202. The function of the learning means 203 is the same as the function of the learning means 104 of the first embodiment.

The learning data selection means 204 is a function of selecting learning data including feature amount data extracted from a layer lower than the layer to be executed fine tuning from the learning data stored in the learning data storage means 202. Have.

The re-learning range determining means 205 has a function of determining one or more candidates for a range (re-learning range) of a layer to be executed for fine tuning. The re-learning range determining means 205 determines one or a plurality of candidates for the range of the layer to be executed fine tuning from the layers constituting the neural network type classifier 201.

For example, re-learning range determining unit 205, "higher layers than the L ₁ layer" candidates ranging executed layers of fine-tuned to the neural network discriminator shown in FIG. 11, "the L ₂ Top layers "than the layer, to determine three ways" layer higher than the L ₃ layer ". The re-learning range determining means 205 outputs information indicating a candidate range of the determined layer to be executed for fine tuning.

The learning data selection means 204 pays attention to one piece of information indicating a candidate of the range of the layer to be executed fine tuning determined by the re-learning range determination means 205. Next, the learning data selection means 204 selects only the learning data obtained from the learning data stored in the learning data storage means 202 other than the candidates in the range of the layer to be executed for fine tuning.

The learning data selected by the learning data selection means 204 is continuously available data. In general, it is considered preferable that the amount of continuously available data is large. Hereinafter, a specific selection method of the learning data selection means 204 will be described with reference to FIG. FIG. 4 is an explanatory diagram showing an example of learning data stored in the learning data storage means 202.

FIG. 4 shows the data name of the learning data and the content of the data together. The content of the data indicates that the training data includes either image data or feature data. When the training data includes feature data, the content of the data also indicates the layer of the neural network type classifier 201 from which the feature data has been extracted.

FIG. 4 shows that the training data A and the training data B include image data. Further, FIG. 4 shows the learning data C and the training data D includes the feature amount data extracted from the intermediate layer L _3.

Further, FIG. 4 shows that the training data E and the training data F include the feature amount data extracted from the _{intermediate layer L 2.} Also, Figure 4 shows that the training data G and learning data H includes the feature amount data extracted from the intermediate layer L _1.

For example, when the re-learning range determining means 205 determines all the layers including the input layer as candidates for the range to be executed for fine tuning, the learning data selecting means 204 sets the learning data A and the learning data B for fine tuning. Select for the training data to be used.

Further, when the re-learning range determining means 205 determines the intermediate layer L ₃ or higher as a candidate for the range to be executed by the fine tuning, the learning data selecting means 204 sets the learning data A, the learning data B, and the learning data C. And the training data D is selected as the training data to be used for fine tuning.

FIG. 5 is an explanatory diagram showing an example of the relationship between the range of the layer to be executed fine tuning and the amount of training data. As shown in FIG. 5, in general, the narrower the range of the layer to be executed for fine tuning, the larger the amount of learning data used for learning and the amount of learning data that can be reused after fine tuning.

Note that FIG. 5 qualitatively shows changes in the amount of learning data used for learning and the amount of reusable learning data. The relationship between the range of layers to be executed for fine tuning, the amount of training data used for training, and the amount of training data that can be reused after fine tuning is not necessarily represented by a linear function as shown in FIG. is not it. The relationship between the two depends on the amount of image data included in the training data amount and the ratio of the amount of feature amount data sampled from each layer.

The evaluation means 206 has a function of evaluating the discrimination performance of the neural network type classifier 201 after fine tuning is executed.

The evaluation means 206 uses the neural network type classifier 201 updated by the learning means 203 to obtain, for example, the discrimination performance of the neural network type classifier 201 with respect to the evaluation data. The discrimination performance may be expressed by, for example, a positive recognition rate and a false recognition rate.

The re-learning result storage means 207 has a function of temporarily storing information indicating a weight parameter of the neural network type classifier 201 after fine tuning is executed.

The re-learning result storage means 207 stores information indicating the network weight parameter of the neural network type classifier 201 updated by the learning means 203. Further, the re-learning result storage means 207 weights the discrimination performance obtained by the evaluation means 206, the information indicating the candidate of the range of the layer to be executed fine tuning, and the number of training data selected by the learning data selection means 204. Store with information indicating the parameters.

The relearning result selection means 208 selects a relearning result (fine tuning result) based on the discrimination performance evaluated by the evaluation means 206 and the amount of learning data available even after the fine tuning is executed. Has a function. The re-learning result selection means 208 can select the optimum re-learning result based on a unified index in which the discrimination performance and the amount of training data available are taken into consideration.

That is, the re-learning result selection means 208 selects the best neural network type classifier 201. Specifically, the re-learning result selection means 208 has the discrimination performance for each candidate of the range of the layer to be executed fine tuning stored in the re-learning result storage means 207, and the learning selected by the learning data selection means 204. Select based on the number of data.

As the selection criteria of the best neural network type classifier 201, for example, the following criteria can be considered.

For example, when the permissible range of discrimination performance is known, the re-learning result selection means 208 relearns with the smallest number of layers to be fine-tuned from among the re-learning results whose discriminating performance is within the permissible range. Just choose the result. By selecting the re-learning result based on the above criteria, the re-learning result selection means 208 can maximize the amount of reusable learning data.

Further, the re-learning result selection means 208 may calculate a score representing the discrimination performance negatively affected by the decrease in the amount of reusable learning data, for example, as follows.

(Score) = (Discrimination performance) -α × (Reduction amount of learning data) ... Equation (2)

However, the constant α in the equation (2) is a constant for aligning the dimension of the discrimination performance with the dimension of the reduction amount of the learning data. The re-learning result selection means 208 may select the neural network type classifier 201 corresponding to the re-learning result that gives the maximum score.

Further, in order to maximize the amount of reusable learning data, the re-learning result selection means 208 may select the neural network type classifier 201 having the smallest number of layers to be executed for fine tuning. When the criterion for maximizing the amount of reusable learning data is clear from the beginning, the re-learning range determining means 205 may output information indicating only the range of the layer corresponding to the criterion.

In the present embodiment, an example in which a neural network type discriminator for performing pattern recognition processing on a discriminating problem as a subject is used as a neural network type discriminator 201 has been described. However, the neural network type discriminator 201 may be a network type discriminator for solving a problem other than the discrimination problem.

For example, the neural network type classifier 201 may be a network type discriminator trained to perform object detection or a network type discriminator trained to perform a regression problem to predict a numerical value.

When the neural network type classifier 201 is a network type classifier for solving a problem other than the discrimination problem, only the function of the evaluation means 206 and the function of the relearning result selection means 208 are changed.

When the neural network type classifier 201 is a network type classifier learned to perform object detection, the evaluation means 206 may express the detection performance by mAP (mean Average Precision), or the false detection rate and detection. The detection performance may be expressed by the rate.

When the detection performance is expressed by the false detection rate and the detection rate, the evaluation means 206 converts the two-dimensional index of the false detection rate and the detection rate into a one-dimensional index called the detection performance, for example, the detection rate and the false detection. The detection rate and the false detection rate may be added after each rate is weighted.

For example, if the weight attached to the false positive rate is set to a negative value or the absolute value of the weight attached to the false positive rate is made larger than the absolute value of the weight attached to the false positive rate, the evaluation means 206 responds to false positives. The penalty can be taken into consideration more.

Further, when the permissible range of the detection performance is known, the re-learning result selection means 208 relearns with the smallest number of layers to be executed for fine tuning from the re-learning results whose detection performance is within the permissible range. Just choose the result. By selecting the re-learning result based on the above criteria, the re-learning result selection means 208 can maximize the amount of reusable learning data.

Further, the re-learning result selection means 208 may calculate a score representing the detection performance negatively affected by the decrease in the amount of reusable learning data, for example, as follows.

(Score) = (Detection performance) -β × (Reduction amount of learning data) ... Equation (3)

However, the constant β in the equation (3) is a constant for aligning the dimension of the detection performance with the dimension of the reduction amount of the learning data. The re-learning result selection means 208 may select the neural network type classifier 201 corresponding to the re-learning result that gives the maximum score.

Further, in order to maximize the amount of reusable learning data, the re-learning result selection means 208 may select the neural network type classifier 201 having the smallest number of layers to be executed for fine tuning. When the criterion for maximizing the amount of reusable learning data is clear from the beginning, the re-learning range determining means 205 may output information indicating only the range of the layer corresponding to the criterion.

Further, when the neural network type classifier 201 is a network type classifier in which the regression problem is learned, the evaluation means 206 uses, for example, the reciprocal of the squared error between the predicted numerical value and the true value as the performance value of the detection performance. Just ask.

In the above description, a linear expression is shown as an example of an expression in which the re-learning result selection means 208 calculates the score from the value of the discrimination performance and the value of the detection performance. However, when it is required to express the statistical relationship with high accuracy, the re-learning result selection means 208 may use a higher-order calculation formula instead of the linear formula as the formula for calculating the score. .. Further, the re-learning result selection means 208 may calculate the score based on a specific model formula.

[Explanation of operation]
Hereinafter, the operation of executing the fine tuning of the pattern recognition device 200 of the present embodiment will be described with reference to FIG. FIG. 6 is a flowchart showing the operation of the fine tuning execution process by the pattern recognition device 200 of the second embodiment.

First, the re-learning range determining means 205 determines a candidate for the range of the layer to be fine-tuned for the neural network type classifier 201 (step S201). The re-learning range determining means 205 inputs a candidate of the range of the determined layer to be executed fine tuning into the learning data selecting means 204.

Next, the learning data selection means 204 pays attention to one range candidate among the range candidates of the layer to be executed for each fine tuning input in step S201. That is, it enters a re-learning loop (step S202).

The learning data selection means 204 stores the learning data required for performing fine tuning on the candidate of the range of the layer to be executed of the fine tuning of interest in the learning data storage means 202. Select from (step S203). The learning data selection means 204 inputs the selected learning data to the learning means 203.

Next, the learning means 203 performs fine tuning on the neural network type classifier 201 using the learning data selected in step S203 (step S204).

Next, the evaluation means 206 evaluates the discrimination performance of the neural network type classifier 201 with respect to the evaluation data by using the neural network type classifier 201 updated in step S204 (step S205).

Next, the evaluation means 206 stores the discrimination performance evaluated in step S205 in the re-learning result storage means 207 as a re-learning result (step S206).

Further, the learning means 203 sends the information indicating the weight parameter of the neural network type classifier 201 updated in step S204 and the information indicating the candidate of the range of the layer to be executed fine tuning to the re-learning result storage means 207. Include in the stored relearning results. Further, the learning data selection means 204 includes the selected number of learning data in the re-learning result stored in the re-learning result storage means 207.

The pattern recognition device 200 repeatedly executes the processes of steps S203 to S206 until fine tuning is executed for all the candidates of the range of the layer to be executed for each fine tuning determined in step S201. When fine tuning is executed for all the candidates in the range of the layer to be executed for each fine tuning, the pattern recognition device 200 exits the relearning loop (step S207).

Next, the re-learning result selection means 208 selects the optimum re-learning result from the re-learning results stored in the re-learning result storage means 207 (step S208). After selecting the optimum relearning result, the pattern recognition device 200 ends the fine tuning execution process.

Hereinafter, other examples of this embodiment will be described with reference to FIG. 7. The pattern recognition device 200 in the above example includes only one type of neural network type classifier 201.

That is, the feature data extracted from the intermediate layer is only the feature data related to one type of neural network type classifier. In other words, the neural network type classifier from which the feature data was extracted was specified in advance as one type.

However, in the learning data storage means 202 of the present embodiment, each feature amount data extracted from each of a plurality of neural network type classifiers may be stored in a mixed manner. That is, the pattern recognition device 200 may include a plurality of types of neural network type classifiers. When the pattern recognition device 200 includes a plurality of types of neural network type classifiers, each neural network type classifier has network identification information.

In this modification, the neural network type classifier further includes network identification information. The network identification information is information that can distinguish each network model and each network coefficient.

That is, for example, when the neural network type classifiers having the same network structure have different network coefficients, different network identification information is assigned to each neural network type classifier.

The network identification information may be represented by a single numerical value or a character string. Further, the network identification information may be information in which the number of layers, the type of each layer, the filter size, the filter coefficient, the stride width, the activation function, and the like are expanded as a single numerical string according to a predetermined rule.

The types of each layer are a convolution layer, a pooling layer, a full connect layer, a softmax layer, and the like. In addition, each information about the layer such as the filter size may be listed for each layer.

Further, in the training data stored in the training data storage means 202, the position of the layer of the neural network type classifier from which the network identification information capable of identifying the neural network type classifier and the feature amount data included in the training data are extracted. Is associated with the information indicating.

Specifically, regarding the feature amount data among the image data and the feature amount data stored in the learning data storage means 202, the network identification information of the neural network type classifier from which the feature amount data is extracted and the feature amount data are A set with information indicating the position of the extracted layer is stored.

Further, one feature amount data may be associated with a plurality of neural network type classifiers. For example, the pattern recognition device 200 includes a neural network type discriminator A and a neural network type discriminator B having a common network structure from the input layer to the Nth layer and a network weight parameter, although the overall structure is different. Suppose you are.

When the feature amount data extracted from the Nth layer of the neural network type classifier A is stored in the learning data storage means 202, the stored feature amount data includes the Nth layer of the neural network type classifier B. Information indicating that the feature amount data is extracted from is also added. The feature data to which the information is added is used over different fine tunings.

FIG. 7 is an explanatory diagram showing another example of the learning data stored in the learning data storage means 202. FIG. 7 shows the data name of the learning data and the content of the data together.

The content of the data indicates that the training data includes either image data or feature data. When the training data includes feature data, the content of the data also indicates the network identification information of the neural network type classifier from which the feature data is extracted and the layer from which the feature data is extracted.

FIG. 7 shows that the training data A and the training data B include image data. Further, FIG. 7 shows that the training data C and the training data D include the feature amount data extracted from _{the intermediate layer L 3} of the network type classifier identified by the _{network identification information N 1.}

Similarly, the training data C and the training data D are the feature amount data extracted from _{the intermediate layer L 3} of the network type classifier identified by the _{network identification information N 2 and the network type identified by the network identification information N 3.} Contains feature data extracted from the middle layer L _{3 of the classifier.}

That is, the structure of the network type classifier identified by the network identification information N ₁ , the structure of the network type classifier identified by the network identification information N ₂ , and the structure of the network type classifier identified by the network identification information N _3. are common from the input layer to the intermediate layer L _3.

Further, in FIG. 7, the training data E and the training data F are identified by the feature amount data extracted from _{the intermediate layer L 2} of the network type classifier identified by the _{network identification information N 1 and the network identification information N 2.} It is shown that the feature data extracted from _{the intermediate layer L 2} of the network type classifier is included.

That is, the structure of the network type classifier identified by the network identification information N ₁ and the structure of the network type classifier identified by the network identification information N ₂ are common from the input layer to the intermediate layer L _2.

Further, FIG. 7 shows that the learning data G and the learning data H include the feature amount data extracted from _{the intermediate layer L 1} of the network type classifier identified by the _{network identification information N 1.}

The learning data selection means 204 pays attention to one range candidate among the range candidates of the layer to be executed fine tuning determined by the re-learning range determination means 205. Next, the training data selection means 204 selects only the training data obtained from the training data stored in the training data storage means 202 other than the candidates of the range of the layer to be fine-tuned by the neural network type classifier. select.

That is, the learning data selection means 204 of this modification selects only the learning data in which the network identification information indicating the neural network type classifier to be executed fine tuning and the included network identification information are equal.

Therefore, the pattern recognition device 200 of this modification can be used as a plurality of neural network type classifiers within the same range of structure even when various neural network type discriminators are constructed for various shooting points. On the other hand, the same feature data can be used.

Hereinafter, specific examples of the hardware configuration of the pattern recognition device of each embodiment will be described. FIG. 8 is an explanatory diagram showing a hardware configuration example of the pattern recognition device according to the present invention.

The pattern recognition device shown in FIG. 8 includes a CPU (Central Processing Unit) 11, a main storage unit 12, a communication unit 13, and an auxiliary storage unit 14. Further, an input unit 15 for the user to operate and an output unit 16 for presenting the processing result or the progress of the processing content to the user may be provided.

The pattern recognition device shown in FIG. 8 may include a DSP (Digital Signal Processor) instead of the CPU 11. Alternatively, the pattern recognition device shown in FIG. 8 may include the CPU 11 and the DSP together.

The main storage unit 12 is used as a data work area or a data temporary storage area. The main storage unit 12 is, for example, a RAM (Random Access Memory).

The communication unit 13 has a function of inputting and outputting data to and from peripheral devices via a wired network or a wireless network (information communication network).

The auxiliary storage unit 14 is a non-temporary tangible storage medium. Examples of non-temporary tangible storage media include magnetic disks, opto-magnetic disks, CD-ROMs (Compact Disk Read Only Memory), DVD-ROMs (Digital Versatile Disk Read Only Memory), and semiconductor memories.

The input unit 15 has a function of inputting data and processing instructions. The input unit 15 is an input device such as a keyboard or a mouse.

The output unit 16 has a function of outputting data. The output unit 16 is, for example, a display device such as a liquid crystal display device or a printing device such as a printer.

Further, as shown in FIG. 8, in the pattern recognition device, each component is connected to the system bus 17.

The auxiliary storage unit 14 stores, for example, a program for realizing the learning means 104, the learning means 203, the learning data selection means 204, the re-learning range determination means 205, the evaluation means 206, and the re-learning result selection means 208. There is.

The pattern recognition device may be realized by hardware. For example, the pattern recognition device may be equipped with a circuit including hardware components such as an LSI (Large Scale Integration) in which a program for realizing a function as shown in FIG. 1 is incorporated.

Further, the pattern recognition device may be realized by software by executing a program in which the CPU 11 shown in FIG. 8 provides the functions of each component.

When realized by software, each function is realized by software by the CPU 11 loading and executing the program stored in the auxiliary storage unit 14 into the main storage unit 12 and controlling the operation of the pattern recognition device. NS. Further, the CPU 11 may load the learning data or the like stored in the auxiliary storage unit 14 into the main storage unit 12.

Further, a part or all of each component may be realized by a general-purpose circuit (circuitry), a dedicated circuit, a processor, or a combination thereof. These may be composed of a single chip or may be composed of a plurality of chips connected via a bus. A part or all of each component may be realized by a combination of the above-mentioned circuit or the like and a program.

When a part or all of each component is realized by a plurality of information processing devices and circuits, the plurality of information processing devices and circuits may be centrally arranged or distributed. For example, the information processing device, the circuit, and the like may be realized as a form in which each is connected via a communication network, such as a client-and-server system and a cloud computing system.

Next, the outline of the present invention will be described. FIG. 9 is a block diagram showing an outline of the pattern recognition device according to the present invention. The pattern recognition device 20 according to the present invention includes feature amount data output by one of a plurality of layers of a neural network type classifier in which a plurality of layers into which image data for learning is input are connected in a layered manner. The second training including the first training data and the feature amount data output by one layer different from one of the plurality of layers of the neural network type classifier in which image data for learning different from the image data is input. Prediction means 21 (for example, learning means 104) that predicts the discrimination performance of the neural network type classifier after learning using data, and the learning target of the neural network type classifier based on the predicted discrimination performance. It is provided with a determination means 22 (for example, a learning means 104) for determining the range of the layers.

With such a configuration, the pattern recognition device can perform fine tuning while considering the amount of reusable feature data.

Further, the pattern recognition device 20 may include learning means (for example, learning means 104) for learning the neural network type classifier using the feature amount data output by the layers in the determined range.

With such a configuration, the pattern recognition device can perform fine tuning while considering the amount of reusable feature data.

Further, the first training data is a correct answer class corresponding to the image data for learning which is the generation source of the feature amount data included in the first training data, and one output of the feature amount data included in the first training data. The second training data includes the information indicating the layer, the correct answer class corresponding to the image data for learning which is the generation source of the feature amount data included in the second training data, and the feature amount included in the second training data. It may include information indicating one layer that outputs data.

With such a configuration, the pattern recognition device can further improve the accuracy of fine tuning.

Further, the learning means includes a learning image data different from the learning image data that is a generation source of the feature amount data included in the first learning data, and a correct answer class corresponding to the different learning image data. 3 The learning data and the first learning data may be used for learning.

With such a configuration, the pattern recognition device can perform a wider variety of fine tuning.

Further, FIG. 10 is a block diagram showing another outline of the pattern recognition device according to the present invention. The pattern recognition device 30 according to the present invention includes a determination means 31 (for example, a relearning range determination means 205) for determining a candidate for a range of layers to be learned in a neural network type classifier in which a plurality of layers are connected in layers. Learning means 32 (for example,) for learning a neural network type classifier using learning data including feature amount data output by a layer of candidates in a determined range of the neural network type classifier to which image data for training is input. The parameters of the learning means 203), the evaluation means 33 (for example, the evaluation means 206) for evaluating the discrimination performance of the neural network type classifier after learning, and the learned and derived neural network type classifier are determined. A storage means 34 (for example, a re-learning result storage means 207) that stores with a range of candidates, and a selection that selects a parameter from the storage means 34 based on the evaluated discrimination performance and the number of learning data used for learning. Means 35 (for example, re-learning result selection means 208) are provided.

With such a configuration, the pattern recognition device can perform fine tuning while considering the amount of reusable feature data.

Further, the learning data may include information indicating the layer that outputs the feature amount data included in the learning data.

With such a configuration, the pattern recognition device can recognize all the feature amount data extracted from the layer higher than the predetermined layer.

Further, the training data may include the image data for learning which is the generation source of the feature amount data included in the training data, and the feature amount data and the correct answer class corresponding to the image data.

With such a configuration, the pattern recognition device can further improve the accuracy of fine tuning.

Further, the training data may include network information indicating a neural network type classifier that outputs feature data included in the training data.

With such a configuration, the pattern recognition device can support a plurality of types of neural network type classifiers.

Further, the learning means 32 may learn using learning data including network information indicating the neural network type classifier to be learned.

With such a configuration, the pattern recognition device can support a plurality of types of neural network type classifiers.

Further, the storage means 34 may store the parameters, the discrimination performance of the neural network type classifier having the parameters, and the number of learning data used for the learning from which the parameters are derived.

With such a configuration, the pattern recognition device can collectively manage the results of a plurality of fine tunings.

Although the present invention has been described above with reference to the embodiments and examples, the present invention is not limited to the above embodiments and examples. Various changes that can be understood by those skilled in the art can be made within the scope of the present invention in terms of the structure and details of the present invention.

In addition, some or all of the above embodiments may be described as in the following appendix, but are not limited to the following.

(Appendix 1) First training data including feature data output by one of the plurality of layers of a neural network type classifier in which a plurality of layers into which image data for training is input are connected in a layered manner. And the second learning including the feature amount data output by one layer different from the one layer among the plurality of layers of the neural network type classifier to which the image data for learning different from the image data is input. A predictive means for predicting the discrimination performance of the neural network type classifier after learning using data, and a range of layers to be learned of the neural network type classifier are determined based on the predicted discrimination performance. A pattern recognition device comprising a determination means for determining the data.

(Appendix 2) The pattern recognition device according to Appendix 1, further comprising a learning means for learning a neural network type classifier using feature data output by layers in a determined range.

(Appendix 3) The first training data includes a correct answer class corresponding to the image data for learning which is a generation source of the feature amount data included in the first learning data, and the feature amount data included in the first learning data. The second learning data includes the information indicating one output layer, the correct answer class corresponding to the image data for learning which is the generation source of the feature amount data included in the second learning data, and the second learning. The pattern recognition device according to Appendix 2, which includes information indicating one layer that outputs feature amount data included in the data.

(Appendix 4) The learning means includes a learning image data different from the learning image data that is the source of the feature amount data included in the first learning data, and a correct answer class corresponding to the different learning image data. The pattern recognition device according to Appendix 3, wherein the third learning data including the above and the first learning data are used for learning.

(Appendix 5) A determination means for determining a candidate for a range of layers to be learned of a neural network type classifier in which a plurality of layers are connected in layers, and the neural network type classifier to which image data for learning is input. Evaluate the learning means for learning the neural network type classifier using the training data including the feature amount data output by the candidate layer in the determined range, and the discrimination performance of the neural network type classifier after learning. Evaluation means to be used, storage means to store the learned and derived parameters of the neural network type classifier together with the candidates in the determined range, and the evaluated discrimination performance and the number of training data used for learning. A pattern recognition device including a selection means for selecting a parameter from the storage means based on the above.

(Appendix 6) The pattern recognition device according to Appendix 5, wherein the learning data includes information indicating a layer that outputs the feature amount data included in the learning data.

(Appendix 7) The training data includes the image data for learning, which is the generation source of the feature data included in the training data, and the feature data and the correct answer class corresponding to the image data. The pattern recognition device described.

(Supplementary note 8) The pattern recognition device according to any one of Supplementary note 5 to Supplementary note 7, wherein the training data includes network information indicating a neural network type classifier that outputs feature amount data included in the training data.

(Appendix 9) The pattern recognition device according to Appendix 8, wherein the learning means is learning using learning data including network information indicating a neural network type classifier to be learned.

(Appendix 10) The storage means stores the parameters, the discrimination performance of the neural network type classifier having the parameters, and the number of learning data used for learning from which the parameters are derived. The pattern recognition device according to any one of 9.

(Appendix 11) First training data including feature data output by one of the plurality of layers of a neural network type classifier in which a plurality of layers into which image data for training is input are connected in a layered manner. And the second learning including the feature amount data output by one layer different from the one layer among the plurality of layers of the neural network type classifier to which the image data for learning different from the image data is input. Predicting the discriminant performance of the neural network type discriminator after learning using the data, and determining the range of the layer to be trained of the neural network type discriminator based on the predicted discriminant performance. Characteristic pattern recognition method.

(Appendix 12) A candidate for a range of layers to be learned of a neural network type classifier in which a plurality of layers are connected in layers is determined, and the neural network type classifier to which image data for learning is input is determined. The neural network type discriminator is trained using the training data including the feature amount data output by the candidate layer of the range, the discrimination performance of the neural network type discriminator after the training is evaluated, and the training is derived. The parameters of the neural network type classifier are stored in the storage means together with the candidates in the determined range, and the parameters are stored from the storage means based on the evaluated discrimination performance and the number of training data used for learning. A pattern recognition method characterized by selection.

(Appendix 13) A th-th. Includes one training data and feature amount data output by one layer different from the one layer among the plurality of layers of the neural network type classifier to which image data for learning different from the image data is input. Prediction processing that predicts the discrimination performance of the neural network type classifier after learning using the second training data, and the layer to be learned of the neural network type classifier based on the predicted discrimination performance. A pattern recognition program for executing a decision process that determines a range.

(Appendix 14) A decision process for determining a candidate for a range of layers to be learned by a neural network type classifier in which a plurality of layers are connected in layers, and the neural network type identification in which image data for learning is input to the computer. The learning process for learning the neural network type classifier using the training data including the feature amount data output by the candidate layer in the determined range of the device, and the discrimination performance of the neural network type classifier after learning. Evaluation processing to be evaluated, storage processing to store the learned and derived parameters of the neural network type classifier in the storage means together with the candidates in the determined range, and training data used for the evaluated discrimination performance and learning. A pattern recognition program for executing a selection process for selecting a parameter from the storage means based on the number of.

11 CPU
12 Main storage unit 13 Communication unit 14 Auxiliary storage unit 15 Input unit 16 Output unit 17 System bus 20, 30, 100, 200 Pattern recognition device 21 Prediction means 22, 31 Determining means 32, 104, 203 Learning means 33, 206 Evaluation means 34 Storage means 35 Selection means 101, 201 Neural network type classifier 102 First learning data storage means 103 Second learning data storage means 202 Learning data storage means 204 Learning data selection means 205 Re-learning range determination means 207 Re-learning result storage means 208 Relearning result selection means

The present invention relates to a pattern recognition apparatus and a pattern recognition how, a pattern recognition apparatus and a pattern recognition how in particular applying statistical pattern recognition techniques.

The present invention solves the above problems, and an object thereof is to provide a pattern recognition apparatus and a pattern recognition how can perform fine tuning taking into account the amount of reusable feature data.

The pattern recognition device according to the present invention is a neural network type discriminator in which a plurality of layers are connected in a layered manner. The learning means for learning the neural network type classifier using the training data including the feature amount data output by the candidate layer in the determined range of the classifier, and the discriminating performance of the neural network type classifier after learning. Evaluation means to evaluate, storage means to store the parameters of the trained and derived neural network type classifier together with the candidates of the determined range, and the evaluated discrimination performance and the number of training data used for training. The training data includes a selection means for selecting a parameter from the storage means based on the training data, and the training data includes network information indicating a neural network type classifier that outputs the feature amount data included in the training data .

Claims (14)

The first training data including the feature amount data output by one of the plurality of layers of the neural network type classifier in which a plurality of layers into which the image data for training is input are connected in a layered manner, and the image. Used by the second learning data including the feature amount data output by one layer different from the one layer among the plurality of layers of the neural network type classifier to which image data for learning different from the data is input. A predictive means for predicting the discrimination performance of the neural network type classifier after being trained and
A pattern recognition device including a determination means for determining a range of layers to be learned of the neural network type classifier based on predicted discrimination performance. The pattern recognition device according to claim 1, further comprising a learning means for learning a neural network type classifier using feature data output by layers in a determined range. The first learning data includes a correct answer class corresponding to the image data for learning that is the generation source of the feature amount data included in the first learning data, and one that outputs the feature amount data included in the first learning data. Including information indicating the layer
The second training data is a correct answer class corresponding to the image data for learning that is the generation source of the feature data included in the second training data, and one that outputs the feature data included in the second training data. The pattern recognition device according to claim 2, which includes information indicating a layer. The learning means includes a third learning image data different from the learning image data that is the generation source of the feature amount data included in the first learning data, and a correct answer class corresponding to the different learning image data. The pattern recognition device according to claim 3, wherein the learning data and the first learning data are used for learning. A determinant that determines a candidate for the range of layers to be trained in a neural network type classifier in which multiple layers are connected in layers,
A learning means for learning the neural network type classifier using learning data including feature data output by a layer of candidates in a determined range of the neural network type classifier to which image data for training is input, and a learning means for learning the neural network type classifier.
An evaluation means for evaluating the discrimination performance of the neural network type classifier after learning, and
A storage means for storing the learned and derived parameters of the neural network type classifier together with the candidates in the determined range, and
A pattern recognition device including a selection means for selecting a parameter from the storage means based on the evaluated discrimination performance and the number of learning data used for learning. The pattern recognition device according to claim 5, wherein the learning data includes information indicating a layer that outputs feature amount data included in the learning data. The fifth or sixth aspect of claim 5 or 6, wherein the learning data includes an image data for learning which is a generation source of the feature amount data included in the learning data, the feature amount data, and a correct answer class corresponding to the image data. Pattern recognition device. The pattern recognition device according to any one of claims 5 to 7, wherein the learning data includes network information indicating a neural network type classifier that outputs feature data included in the learning data. The pattern recognition device according to claim 8, wherein the learning means is learning using learning data including network information indicating a neural network type classifier to be learned. The storage means according to claims 5 to 9 store the parameters, the discrimination performance of the neural network type classifier having the parameters, and the number of learning data used for learning from which the parameters are derived. The pattern recognition device according to any one of the items. The first training data including the feature amount data output by one of the plurality of layers of the neural network type classifier in which a plurality of layers into which the image data for training is input are connected in a layered manner, and the image. Used by the second learning data including the feature amount data output by one layer different from the one layer among the plurality of layers of the neural network type classifier to which image data for learning different from the data is input. Predicting the discrimination performance of the neural network type classifier after being learned and learned,
A pattern recognition method characterized in that the range of layers to be learned of the neural network type classifier is determined based on the predicted discrimination performance. A neural network type classifier in which multiple layers are connected in layers determines a candidate for the range of layers to be trained.
The neural network type classifier is trained using the training data including the feature amount data output by the candidate layer in the determined range of the neural network type classifier to which the image data for training is input.
After being trained, the discrimination performance of the neural network type classifier was evaluated, and
The learned and derived parameters of the neural network type classifier are stored in the storage means together with the candidates in the determined range, and the parameters are stored in the storage means.
A pattern recognition method characterized in that parameters are selected from the storage means based on the evaluated discrimination performance and the number of learning data used for learning. On the computer
The first training data including the feature amount data output by one of the plurality of layers of the neural network type classifier in which a plurality of layers into which image data for training is input are connected in a layered manner, and the image. Used by the second learning data including the feature amount data output by one layer different from the one layer among the plurality of layers of the neural network type classifier to which image data for learning different from the data is input. A prediction process for predicting the discrimination performance of the neural network type classifier after being trained, and a determination process for determining the range of layers to be learned of the neural network type classifier based on the predicted discrimination performance. A pattern recognition program for execution. On the computer
A decision process that determines a candidate for the range of layers to be trained in a neural network type classifier in which multiple layers are connected in layers.
A learning process for learning the neural network type classifier using learning data including feature data output by a layer of candidates in a determined range of the neural network type classifier to which image data for training is input.
Evaluation processing for evaluating the discrimination performance of the neural network type classifier after learning,
Based on the storage process of storing the learned and derived parameters of the neural network type classifier in the storage means together with the candidates of the determined range, and the evaluated discrimination performance and the number of training data used for learning. A pattern recognition program for executing a selection process for selecting a parameter from the storage means.

Images