Image feature learning device, image feature learning method, image feature extraction device, image feature extraction method, and program
The purpose of the present invention is to enable learning of a neural network for extracting features of images having high robustness from an undiscriminating image region while minimizing the number of parameters of a pooling layer. A parameter learning unit 130 learns parameters of each layer in a convolutional neural network configured by including a fully convolutional layer for performing convolution of an input image to output a feature tensor of the input image, a weighting matrix estimation layer for estimating a weighting matrix indicating a weighting of each element of the feature tensor, and a pooling layer for extracting a feature vector of the input image based on the feature tensor and the weighting matrix. The parameter learning unit 130 learns the parameters such that a loss function value obtained by calculating a loss function expressed by using a distance between a first feature vector of a first image and a second feature vector of a second image, which are relevant images and are obtained by applying the convolutional neural network, becomes smaller.
1. An image feature learning device for learning a convolutional neural network for extracting a feature vector of an input image based on training data as a set of a first image, a second image, and a label indicating whether the first image and the second image are relevant images, comprising:
an input receiver configured to receive input of the training data, the convolutional neural network being configured by including:
a fully convolutional layer for performing convolution of the input image to output a feature tensor of the input image;
a weighting matrix estimation layer for estimating a weighting matrix indicating a weighting of each element of the feature tensor; and
a pooling layer for extracting a feature vector of the input image based on the feature tensor and the weighting matrix; and
a parameter learner configured to learn parameters of each layer in the convolutional neural network such that a loss function value obtained by calculating,
based on:
a first feature vector obtained by applying the first image included in the training data to the convolutional neural network as the input image,
a second feature vector obtained by applying the second image included in the training data to the convolutional neural network as the input image, and
the label included in the training data,
a loss function expressed by using a distance between the first feature vector of the first image and the second feature vector of the second image, which are relevant images, becomes smaller.
2. The image feature learning device according to claim 1 ,
wherein the feature tensor is a three-dimensional tensor corresponding to a height, a width, and channels of the input image,
wherein the weighting matrix estimation layer inputs the feature tensor to estimate the weighting matrix indicating a weighting of each two-dimensional element corresponding to the height and the width of the input image, and
wherein the pooling layer calculates, for each channel of the input image, a weighted generalized mean of two-dimensional elements corresponding to the height and the width for the channel among three-dimensional elements of the feature tensor and an element of the weighting matrix and extracts the feature vector formed of a value calculated for each channel.
3. The image feature learning device according to claim 2 , further comprising:
generating a plurality of the first images and a plurality of the second images by converting, for each of a plurality of magnifications, the first image and the second image included in the training data to a size enlarged or reduced by using the magnification,
wherein the image feature learning device learns parameters of each layer in the convolutional neural network such that: a loss function value obtained by calculating,
based on:
a first integrated feature vector obtained by integrating a plurality of the first feature vectors obtained by applying each of a plurality of the first images generated by the size image generator to the convolutional neural network as the input image,
a second integrated feature vector obtained by integrating a plurality of the second feature vectors obtained by applying each of a plurality of the second images generated by the size image generator to the convolutional neural network as the input image, and
the label included in the training data,
the loss function expressed by using a distance between the first integrated feature vector of the first image and the second integrated feature vector of the second image, which are relevant images, becomes smaller.
4. The image feature learning device according to claim 2 , the device further comprising:
the input receiver configured to receive an input image; and
an image feature extractor configured to extract a feature vector of the input image by using the convolutional neural network learned by the image feature learning device.
5. The image feature learning device according to claim 1 , further comprising:
a size image generator configured to generate a plurality of the first images and a plurality of the second images by converting, for each of a plurality of magnifications, the first image and the second image included in the training data to a size enlarged or reduced by using the magnification, wherein the image feature learning device learns parameters of each layer in the convolutional neural network such that:
a loss function value obtained by calculating,
based on:
a first integrated feature vector obtained by integrating a plurality of the first feature vectors obtained by applying each of a plurality of the first images generated by the size image generator to the convolutional neural network as the input image,
a second integrated feature vector obtained by integrating a plurality of the second feature vectors obtained by applying each of a plurality of the second images generated by the size image generator to the convolutional neural network as the input image, and
the label included in the training data,
the loss function expressed by using a distance between the first integrated feature vector of the first image and the second integrated feature vector of the second image, which are relevant images, becomes smaller.
6. The image feature learning device according to claim 1 , the device further comprising:
the input receiver configured to receive an input image; and
an image feature extractor configured to extract a feature vector of the input image by using the convolutional neural network learned by the image feature learning device.
7. The image feature learning device according to claim 6 , further comprising a size image generator configured to generate a plurality of the input images by converting, for each of a plurality of magnifications, the input image to a size enlarged or reduced by using the magnification, wherein the image feature extractor configured to extract, for each of the plurality of the input images, a feature vector by using the convolutional neural network, and set a vector obtained by integrating feature vectors extracted for the plurality of the input images as a feature vector of the input image.
8. An image feature learning method for learning a convolutional neural network for extracting a feature vector of an input image based on training data as a set of a first image, a second image, and a label indicating whether the first image and the second image are relevant images, comprising:
receiving, by a receiver, input of the training data, the convolutional neural network being configured by including:
a fully convolutional layer for performing convolution of the input image to output a feature tensor of the input image;
a weighting matrix estimation layer for estimating a weighting matrix indicating a weighting of each element of the feature tensor; and
a pooling layer for extracting a feature vector of the input image based on the feature tensor and the weighting matrix; and
learning, by a parameter learner, parameters of each layer in the convolutional neural network such that a loss function value obtained by calculating,
based on:
a first feature vector obtained by applying the first image included in the training data to the convolutional neural network as the input image,
a second feature vector obtained by applying the second image included in the training data to the convolutional neural network as the input image, and
the label included in the training data,
a loss function expressed by using a distance between the first feature vector of the first image and the second feature vector of the second image, which are relevant images, becomes smaller.
9. The image feature learning method of claim 8 , comprising:
receiving, by the receiver, an input image; and
extracting, by an image feature extractor, a feature vector of the input image by using the convolutional neural network learned by the image feature learning method according to claim 8 .
10. The image feature learning method of claim 8 ,
wherein the feature tensor is a three-dimensional tensor corresponding to a height, a width, and channels of the input image,
wherein the weighting matrix estimation layer inputs the feature tensor to estimate the weighting matrix indicating a weighting of each two-dimensional element corresponding to the height and the width of the input image, and
wherein the pooling layer calculates, for each channel of the input image, a weighted generalized mean of two-dimensional elements corresponding to the height and the width for the channel among three-dimensional elements of the feature tensor and an element of the weighting matrix and extracts the feature vector formed of a value calculated for each channel.
11. The image feature learning method of claim 10 , further comprising:
generating a plurality of the first images and a plurality of the second images by converting, for each of a plurality of magnifications, the first image and the second image included in the training data to a size enlarged or reduced by using the magnification, wherein the image feature learning device learns parameters of each layer in the convolutional neural network such that:
a loss function value obtained by calculating,
based on:
a first integrated feature vector obtained by integrating a plurality of the first feature vectors obtained by applying each of a plurality of the first images generated by the size image generator to the convolutional neural network as the input image,
a second integrated feature vector obtained by integrating a plurality of the second feature vectors obtained by applying each of a plurality of the second images generated by the size image generator to the convolutional neural network as the input image, and
the label included in the training data,
the loss function expressed by using a distance between the first integrated feature vector of the first image and the second integrated feature vector of the second image, which are relevant images, becomes smaller.
12. The image feature learning method of claim 8 , further comprising:
generating a plurality of the first images and a plurality of the second images by converting, for each of a plurality of magnifications, the first image and the second image included in the training data to a size enlarged or reduced by using the magnification, wherein the image feature learning device learns parameters of each layer in the convolutional neural network such that:
a loss function value obtained by calculating,
based on:
a first integrated feature vector obtained by integrating a plurality of the first feature vectors obtained by applying each of a plurality of the first images generated by the size image generator to the convolutional neural network as the input image,
a second integrated feature vector obtained by integrating a plurality of the second feature vectors obtained by applying each of a plurality of the second images generated by the size image generator to the convolutional neural network as the input image, and
the label included in the training data,
the loss function expressed by using a distance between the first integrated feature vector of the first image and the second integrated feature vector of the second image, which are relevant images, becomes smaller.
13. The image feature learning method of claim 8 , further comprising:
receiving an input image; and
extracting a feature vector of the input image by using the learnt convolutional neural network.
14. The image feature learning method of claim 8 , further comprising:
generating a plurality of the input images by converting, for each of a plurality of magnifications, the input image to a size enlarged or reduced by using the magnification,
extracting, for each of the plurality of the input images, a feature vector by using the convolutional neural network, and
setting a vector obtained by integrating feature vectors extracted for the plurality of the input images as a feature vector of the input image.
15. A computer-readable non-transitory recording medium storing a computer-executable program for image feature learning processing for learning a convolutional neural network for extracting a feature vector of an input image based on training data as a set of a first image, a second image, and a label indicating whether the first image and the second image are relevant images, the computer-executable program when executed by a processor causes a computer system to:
receive input of the training data, the convolutional neural network being configured by including:
a fully convolutional layer for performing convolution of the input image to output a feature tensor of the input image;
a weighting matrix estimation layer for estimating a weighting matrix indicating a weighting of each element of the feature tensor; and
a pooling layer for extracting a feature vector of the input image based on the feature tensor and the weighting matrix; and
learn parameters of each layer in the convolutional neural network such that a loss function value obtained by calculating,
based on:
a first feature vector obtained by applying the first image included in the training data to the convolutional neural network as the input image,
a second feature vector obtained by applying the second image included in the training data to the convolutional neural network as the input image, and
the label included in the training data,
a loss function expressed by using a distance between the first feature vector of the first image and the second feature vector of the second image, which are relevant images, becomes smaller.
16. The computer-readable non-transitory recording medium of claim 15 , wherein the feature tensor is a three-dimensional tensor corresponding to a height, a width, and channels of the input image,
wherein the weighting matrix estimation layer inputs the feature tensor to estimate the weighting matrix indicating a weighting of each two-dimensional element corresponding to the height and the width of the input image, and
wherein the pooling layer calculates, for each channel of the input image, a weighted generalized mean of two-dimensional elements corresponding to the height and the width for the channel among three-dimensional elements of the feature tensor and an element of the weighting matrix and extracts the feature vector formed of a value calculated for each channel.
17. The computer-readable non-transitory recording medium of claim 16 , the computer-executable program when executed by the processor further causes the computer system to:
generate a plurality of the first images and a plurality of the second images by converting, for each of a plurality of magnifications, the first image and the second image included in the training data to a size enlarged or reduced by using the magnification, wherein the image feature learning device learns parameters of each layer in the convolutional neural network such that:
a loss function value obtained by calculating,
based on:
a first integrated feature vector obtained by integrating a plurality of the first feature vectors obtained by applying each of a plurality of the first images generated by the size image generator to the convolutional neural network as the input image,
a second integrated feature vector obtained by integrating a plurality of the second feature vectors obtained by applying each of a plurality of the second images generated by the size image generator to the convolutional neural network as the input image, and
the label included in the training data,
the loss function expressed by using a distance between the first integrated feature vector of the first image and the second integrated feature vector of the second image, which are relevant images, becomes smaller.
18. The computer-readable non-transitory recording medium of claim 15 , the computer-executable program when executed by the processor further causes the computer system to:
generate a plurality of the first images and a plurality of the second images by converting, for each of a plurality of magnifications, the first image and the second image included in the training data to a size enlarged or reduced by using the magnification, wherein the image feature learning device learns parameters of each layer in the convolutional neural network such that:
a loss function value obtained by calculating,
based on:
a first integrated feature vector obtained by integrating a plurality of the first feature vectors obtained by applying each of a plurality of the first images generated by the size image generator to the convolutional neural network as the input image,
a second integrated feature vector obtained by integrating a plurality of the second feature vectors obtained by applying each of a plurality of the second images generated by the size image generator to the convolutional neural network as the input image, and
the label included in the training data,
the loss function expressed by using a distance between the first integrated feature vector of the first image and the second integrated feature vector of the second image, which are relevant images, becomes smaller.
19. The computer-readable non-transitory recording medium of claim 15 , the computer-executable program when executed by the processor further causes the computer system to:
receive an input image; and
extract a feature vector of the input image by using the convolutional neural network learned by the image feature learning device.
20. The computer-readable non-transitory recording medium of claim 15 , the computer-executable program when executed by the processor further causes the computer system to:
generate a plurality of the input images by converting, for each of a plurality of magnifications, the input image to a size enlarged or reduced by using the magnification;
extract, for each of the plurality of the input images, a feature vector by using the convolutional neural network, and
set a vector obtained by integrating feature vectors extracted for the plurality of the input images as a feature vector of the input image.