Learning device, noise removal device, and program
A learner, a noise reduction device, and a program that can train a noise reducer even in a case where learning data having a smaller amount of noise than a normal captured image cannot be acquired or in a case where an amount of noise in learning data having the same noise as the normal captured image is unknown are provided. A learner includes a noise reducer that reduces noise of first learning data using a noise reducer that is trainable, a noise addition unit that adds predetermined noise to be added to the first learning data in which the noise is reduced, and a learning unit that trains the noise reducer to bring a distribution of the first learning data to which the noise to be added is added close to a distribution of second learning data.
1 . A learner comprising:
at least one processor,
wherein the processor is configured to:
reduce original noise of first learning data which is image data acquired by an imaging apparatus that includes the original noise generated in the imaging apparatus, using a noise reducer that is trainable;
add simulated noise to the first learning data in which the original noise is reduced to generate simulated-noise added first learning data, the simulated noise being generated, by a noise generator, to simulate the original noise generated in the imaging apparatus; and
train the noise reducer to minimize deviation between a distribution of the simulated-noise added first learning data and a distribution of second learning data that is the same data as the first learning data, and which includes the original noise of the first learning data,
wherein the learner is configured to train the noise reducer, and
wherein the noise generator comprises a plurality of noise generators respectively corresponding to a plurality of types of imaging apparatus or image data, and the noise generator is configured to select one of the plurality of noise generators based on a type of the imaging apparatus or the image data associated with the first learning data.
2 . The learner according to claim 1 ,
wherein the noise reducer is a deep learning model.
3 . The learner according to claim 1 ,
wherein target data for reducing noise via the noise reducer is image data.
4 . The learner according to claim 3 ,
wherein the image data is a radiation image.
5 . The learner according to claim 1 ,
wherein the processor is configured to train the noise reducer to make the distribution of the simulated-noise added first learning data and the distribution of the second learning data unidentifiable by a predetermined identifier.
6 . The learner according to claim 1 ,
wherein the deviation is a Wasserstein distance or a value based on an information amount.
7 . A noise reduction device comprising:
a noise reducer; and
an acquirer,
wherein the noise reducer is trained by a learner that includes:
at least one processor that is configured to:
reduce original noise of first learning data, which is image data acquired by an imaging apparatus that includes the original noise generated in the imaging apparatus, using the noise reducer that is trainable;
add simulated noise to the first learning data in which the original noise is reduced to generate simulated-noise added first learning data, the simulated noise being generated, by a noise generator, to simulate the original noise generated in the imaging apparatus; and
train the noise reducer to minimize deviation between a distribution of the simulated-noise added first learning data and a distribution of second learning data that is the same as the first learning data, and which includes the original noise of the first learning data,
wherein the acquirer acquires the first learning data, in which the original noise is reduced, by inputting the first learning data into the noise reducer, and
wherein the noise generator comprises a plurality of noise generators respectively corresponding to a plurality of types of imaging apparatus or image data, and the noise generator is configured to select one of the plurality of noise generators based on a type of the imaging apparatus or the image data associated with the first learning data.
8 . A non-transitory storage medium storing a learning program causing a computer to execute a learning process comprising:
reducing original noise of first learning data which is image data acquired by an imaging apparatus that includes the original noise generated in the imaging apparatus, using a noise reducer that is trainable;
adding simulated noise to the first learning data in which the original noise is reduced to generate simulated-noise added first learning data, the simulated noise being generated, by a noise generator, to simulate the original noise generated in the imaging apparatus; and
training the noise reducer to minimize deviation between a distribution of the simulated-noise added first learning data and a distribution of second learning data that is the same data as the first learning data, and which includes the original noise of the first learning data,
wherein the learning process is configured to train the noise reducer, and
wherein the noise generator comprises a plurality of noise generators respectively corresponding to a plurality of types of imaging apparatus or image data, and the noise generator is configured to select one of the plurality of noise generators based on a type of the imaging apparatus or the image data associated with the first learning data.
9 . The learner according to claim 1 ,
wherein the learner further includes an identifier configured to receive the first learning data and the simulated-noise added first learning data and to output a scalar value for each of the received data, and
wherein the processor is configured to train the noise reducer such that a Wasserstein distance, which is calculated based on outputs of the identifier, between a distribution of the first learning data and a distribution of the simulated-noise added first learning data is minimized.