Patent Application
20240212323
The present disclosure relates to an image processing technology based on machine learning.
Human beings can learn new knowledge through experiences over a long period of time and can maintain old knowledge without forgetting it. Meanwhile, the knowledge of a convolutional neutral network (CNN) depends on the dataset used in learning. To adapt to a change in data distribution, it is necessary to re-train CNN parameters in response to the entirety of the dataset. In CNN, the precision estimation for old tasks will be decreased as new tasks are learned. Thus, catastrophic forgetting cannot be avoided in CNN. Namely, the result of learning old tasks is forgotten as new tasks are being learned in successive learning.
Incremental learning or continual learning is proposed as a scheme to avoid catastrophic forgetting. Continual learning is a learning method that improves a current trained model to learn new tasks and new data as they occur, instead of training the model from scratch. One scheme of continual learning is known as regularization-based continual learning, in which learning is performed by using regularization loss (PATENT LITERATURE 1).
The technology described in PATENT LITERATURE 1 has a problem in that catastrophic forgetting cannot be sufficiently reduced.
An image processing apparatus according to an embodiment includes: a basic class selection unit that selects, in response to input data, a base class based on an embedding vector output by a basic neural network that has learned the base class and a centroid vector of the base class; a continual learning unit that continually learns an additional class by using an additional neural network that has learned the base class; an additional class selection unit that selects, in response to the input data, an additional class based on an embedding vector output by the additional neural network subjected to continual learning and centroid vectors of the base class and the additional class; and a classification determination unit that classifies the input data based on the base class selected by the base class selection unit and the additional class selected by the additional class selection unit.
Another embodiment of this embodiment relates to an image processing method. The method includes: selecting, in response to input data, a base class based on an embedding vector output by a basic neural network that has learned the base class and a centroid vector of the base class; continually learning an additional class by using an additional neural network that has learned the base class; selecting, in response to the input data, an additional class based on an embedding vector output by the additional neural network subjected to continual learning and centroid vectors of the base class and the additional class; and classifying the input data based on the base class selected by the selecting of a base class and the additional class selected by the selecting of an additional class.
Optional combinations of the aforementioned constituting elements, and implementations of the embodiment in the form of methods, apparatuses, systems, recording mediums, and computer programs may also be practiced as additional modes of the present embodiment.
The disclosure will be described with reference to the following drawings.
The invention will now be described by reference to the preferred embodiments. This does not intend to limit the scope of the present invention, but to exemplify the invention.
In this embodiment, machine learning in which continual learning and metric learning are combined is performed. An image will be described here as an example of input data, but the input data is not limited to an image. Metric learning is known as a method of learning a embedding space (feature space) by considering the relationship between images (see, for example, NON-PATENT LITERATURE 1). Metric learning is used in various fields such as information retrieval, data classification, and image recognition. Continual learning that uses regularization loss for the purpose of learning can be combined with metric learning that uses metric loss.
In this embodiment, class incremental learning, which is a type of continual learning, is used (see, for example, NON-PATENT LITERATURE 2, NON-PATENT LITERATURE 3). NON-PATENT LITERATURE 2 teaches performing class incremental learning in one neural network. NON-PATENT LITERATURE 3 teaches performing class incremental learning in two neural networks having different learning rates to perform classification in a connected feature space in which the feature spaces of the two neural networks are connected.
In this embodiment, the basic neural network that has learned a base class is not changed, and an additional neural network that has learned the base class and continues to learn an additional class is updated. Classification (class selection) is applied to the input image by using the basic neural network and the additional neural network, respectively, and the input image is classified into a class with a higher accuracy (closer distance).
First, a neural network that has learned the base class and the centroid vector of the base class derived by using the neural network are acquired. The neural network that has learned the base class may be acquired from a network, or a neural network may be trained by using a dataset including the base class. It is desirable that the neural network that has learned the base class is not trained for classification learning but trained for metric learning (embedding learning). The centroid vector of the base class may be acquired from the network. Alternatively, an image of the base class may be input to a trained neural network, and the centroid vector of each class may be derived by determining, for each class, the centroid of the embedding vectors output from the trained neural network. The number of centroid vectors per class is assumed to be 1 but may be plural.
The neural network that has learned the base class is set to be the basic neural network processing unit 10 and the additional neural network processing unit 30 (S10).
The centroid vector of the base class derived by using the neural network that has learned the base class is set in the base class selection unit 20 and the additional class selection unit 40 (S20). The base class selection unit 20 and the additional class selection unit 40 store the centroid vector of the base class, respectively.
Next, a learning session i, which is continual learning, is repeated N times (i=1, 2, . . . , N) (S30).
First, the additional neural network processing unit 30 inputs all images of each additional class included in an additional training dataset to the additional neural network prior to the learning session i, thereby deriving embedding vectors in all images of the additional class. The centroid derivation unit 60 derives the centroid vector of the additional class from the embedding vectors of all images of the additional class (S40). The centroid vector of the additional class in this case is the pre-learning centroid vector. It is noted that the centroid vector of the additional class is derived for all additional classes.
The continual learning unit 50 then proceeds to the learning session i and subjects the additional neural network to continual learning by using the additional training dataset including the additional class (S50).
The additional neural network processing unit 30 then inputs all images of each additional class included in the additional training dataset to the additional neural network subjected to the learning session i, thereby deriving embedding vectors in all images of the additional class. The centroid derivation unit 60 derives the centroid vector of the additional class from the embedding vectors of all images of the additional class (S60). The centroid vector of the additional class in this case is the post-learning centroid vector. It is noted that the centroid vector of the additional class is derived for all additional classes.
The additional class selection unit 40 then deletes the centroid vector of the base class stored (S70). It is assumed here that the number of centroid vectors of the base classes deleted is the number of additional classes added in learning session I. It is assumed that the centroid vector of the base class to be deleted is the one nearest the centroid vector of the additional class added in the learning session i. After all the centroid vectors of the selected base classes are deleted, the centroid vectors are not deleted. This ensures that the number of centroid vectors stored by the basic class selection unit 20 and the number of centroid vectors stored by the additional class selection unit 40 are identical.
The centroid vector correction unit 70 then corrects the centroid vector of known classes stored by the additional class selection unit 40 (S80). The known classes include base classes and additional classes in the learning session (i−1). Additional classes in learning session i do not need to be corrected. i is incremented by 1 (S90), control is returned to step S30, steps S40-S80 are repeated until i=N, and the process is ended when i exceeds N.
For correction of the centroid vector of the learned (known) class, the method described with reference to
The centroid vector correction unit 70 corrects the centroid vector of the learned class based on the centroid vector of the class prior to continuation learning and the centroid vector of the class subsequent to continuation learning within a predetermined distance from the centroid vector of the learned class (known class). Specifically, the centroid vector correction unit 70 determines the amount of movement from the centroid vector of the class prior to continual learning to the centroid vector of the class subsequent to continual learning and calculates the average movement amount by referring to the amounts of movement. The centroid vector correction unit 70 corrects the centroid vector of the learned class by adding the average movement amount to the centroid vector of the learned class.
In NON-PATENT LITERATURE 2, a pre-learning embedding vector within the radius R of the centroid vector of a known class is used for correction. This embodiment differs in that both the centroid vector of the class prior to continual learning and the centroid vector of the class subsequent to continual learning are used for correction. Calculation of the average movement amount will be less affected by fine image-to-image fluctuations by using a large number of centroid vectors. In this embodiment, therefore, both the centroid vector of the class prior to continual learning and the centroid vector of the class subsequent to continual learning within a predetermined distance from the centroid vector of the learned class are used for correction.
The configuration and operation of the continual learning unit 50 will be described in further detail.
The basic training dataset is a supervised dataset including a large number of base classes (e.g., about 100 to 1000 classes), wherein each class is comprised of a large number of images (e.g., 3000 images). The basic training dataset is assumed to have a sufficient amount of data to allow learning a general classification task alone.
On the other hand, the additional training dataset is a supervised dataset including a small number of additional classes (e.g., about 2 to 10 classes), wherein each additional class is comprised of a small number of images (e.g., about 1 to 5 images). Training data comprised of a set of three images, including an anchor image belonging to a given class, a positive image belonging to the same class as the anchor image, and a negative image belonging to a class different from that of the anchor image, is input to the neural network that should be trained. The reason that two classes are assumed as a small number of classes is that a class that contains negative images and that is not learned is necessary even if the number of classes learned is one. It is assumed here that the set includes a small number of images but may include a large number of images provided that the number of classes is small.
The continual learning unit 50 calculates the total loss L as given by the following expression by adding the metric loss Lml and the regularization loss Lr. The continual learning unit 50 trains the neural network so as to minimize the overall loss L.
where Σ indicates taking a sum with for an input image.
Triplet loss is used as metric loss. The triplet loss Lml is calculated by the following expression based on the embedding vector of the anchor image, the embedding vector of the positive image, and the embedding vector of the negative image.
Lml=dp−dn+a
where dp denotes the Euclidean distance between the embedding vector of the anchor image and the embedding vector of the positive image. dn denotes the Euclidean distance between the embedding vector of the anchor image and the embedding vector of the negative image. a is the offset.
The regularization loss Lr is an embedding vector loss Lrv as given by the following expression for minimizing the difference between embedding vectors output before and after the learning session when an image is input to the neural network.
where V (i) denotes an embedding vector output by the neural network in the learning session i. V(i−1) is the embedding vector output by the neural network in the learning session (i−1). ||·|| is a symbol that indicates the meaning of calculating a Frobenius norm.
The basic neural network processing unit 10 inputs an image subject to classification to the basic neural network, and the additional neural network processing unit 30 inputs the image subject to classification to the additional neural network subjected to continual learning (S100).
The basic neural network processing unit 10 supplies the embedding vector of the image subject to classification output from the basic neural network to the basic class selection unit 20. The additional neural network processing unit 30 supplies the embedding vector of the image subject to classification output from the additional neural network to the additional class selection unit 40 (S110).
The base class selection unit 20 selects a base class based on the base embedding vector output by the basic neural network (S120). Specifically, a base class having a centroid vector closest to the base embedding vector is selected.
The additional class selection unit 40 selects an additional class based on the additional embedding vector output by the additional neural network (S130). Specifically, an additional class having a centroid vector closest to the additional embedding vector is selected. The additional class selection unit 40 does not select a base class even if the class having a centroid vector closest to the additional embedding vector is a base class.
The classification determination unit 80 compares the base class selected by the base class selection unit 20 and the additional class selected by the additional class selection unit 40 and determines the class characterized by the closest distance between the centroid vector and the embedding vector to be class resulting from classification of the image subject to classification (S140). A reciprocal of the distance between the centroid vector and the embedding vector may be taken and treated as if it indicates a probability. Relative magnitude of the probability may be examined, and the class associated with a higher probability may be determined to be the class resulting from classification. When the distance between the centroid vector and the embedding vector is the same between the selected base class and the selected additional class, the additional class is selected as the class resulting from classification.
A variation of the additional class selection unit 40 and the classification determination unit 80 will be described. Only the operations that differ from those of the embodiment will be described. The additional class selection unit 40 selects the centroid vector having the closest distance to the additional embedding vector regardless of whether the class is a base class or an additional class. When the base class selected by the base class selection unit 20 and the base class selected by the additional class selection unit 40 are different, the classification determination unit 80 selects the base class selected by the base class selection unit 20 as the class resulting from classification. The reason that the classification determination unit 80 selects the base class selected by the basic class selection unit 20 as the class resulting from classification is that the basic neural network has learned more data for the basic class. That is, the classification determination unit 80 selects the classification result of the neural network trained with more data.
The above-described various processes in the image processing apparatus 100 can of course be implemented by hardware-based apparatuses such as a CPU and a memory and can also be implemented by firmware stored in a ROM (read-only memory), a flash memory, etc., or by software on a computer, etc. The firmware program or the software program may be made available on, for example, a computer readable recording medium. Alternatively, the program may be transmitted and received to and from a server via a wired or wireless network. Still alternatively, the program may be transmitted and received in the form of data broadcast over terrestrial or satellite digital broadcast systems.
As described above, according to the image processing apparatus 100 of this embodiment, the basic neural network is not subject to continual learning and so does not forget the basic class. Even if the learning session progresses, therefore, the basic neural network can classify into the base class with a high probability. Since the basic neural network does not continually learn an additional class, the basic neural network cannot select an additional class. However, the additional neural network can learn the base class and the additional class such that the features of both are considered and can select an additional class, by continually learning the additional class in addition to the base class.
According to this embodiment, the result of classification by the basic neural network that does not forget the base class and the result of classification by the additional neural network trained to learn the additional class continually are evaluated, and the result of classification having a higher accuracy is selected. Accordingly, it is possible to mitigate catastrophic forgetting and improve accuracy of classification at the same time.
In the case the additional neural network learns only the additional class, the centroid vector of the additional class is likely to be overfitted because the number of data items in the additional class is small. Also, it is highly likely that the centroid vector is overcorrected. In the training of the additional neural network, therefore, it is possible to prevent the centroid vector of the additional class and correction of the centroid vector from varying significantly through overfitting and to reduce overfitting to the centroid vector of the additional class and correction of the centroid vector, by considering, along with the additional class, the embedding vector output by the basic neural network that has learned the base classes with more data.
Further, by keeping the total number of classes comprising the base classes and the additional classes in the additional class selection unit 40 constant, the embedding spaces of the basic neural network and the additional neural network can be kept at similar densities, and the distance in the embedding space of the basic class selection unit 20 and the additional class selection unit 40 can be dealt with as being similar. It is possible to prevent a bias in class selection from being created between the base class selection unit 20 and the additional class selection unit 40.
Described above is an explanation based on an exemplary embodiment. The embodiment is intended to be illustrative only and it will be understood by those skilled in the art that various modifications to combinations of constituting elements and processes are possible and that such modifications are also within the scope of the present invention.
| Number | Date | Country | Kind |
|---|---|---|---|
| 2021-140819 | Aug 2021 | JP | national |
This application is a continuation of application No. PCT/JP2022/021174, filed on May 24, 2022, and claims the benefit of priority from the prior Japanese Patent Application No. 2021-140819, filed on Aug. 31, 2021, the entire content of which is incorporated herein by reference.
| Number | Date | Country | |
|---|---|---|---|
| Parent | PCT/JP2022/021174 | May 2022 | WO |
| Child | 18588056 | US |
