Determination method and recording medium
A determination method for determining the structure of a convolutional neural network includes acquiring N filters having the weights trained using a training image group as the initial values, where N is a natural number greater than or equal to 1, and increasing the number of the filters from N to M, where M is a natural number greater than or equal to 2 and is greater than N, by adding a filter obtained by performing a transformation used in image processing fields on at least one of the N filters.
1. A determination method for determining a structure of a convolutional neural network, the determination method comprising:
acquiring N filters having weights trained using a training image group as initial values, where N is a natural number greater than or equal to 1; and
splitting at least one of the N filters to increase a number of filters from N to M and to acquire M filters, where M is a natural number greater than or equal to 2 and is greater than N,
wherein the splitting adds at least one filter by performing a transformation used in image processing fields on each of the at least one of the N filters.
2. The determination method according to claim 1 , further comprising:
evaluating an identification performance of the M filters by causing the M filters to learn the weights by using the training image group,
wherein, if the identification performance of the M filters is less than or equal to an identification performance of the N filters, splitting at least one of the M filters.
3. The determination method according to claim 1 , further comprising:
merging the M filters into L filters, where L is a natural number greater than or equal to 1 and is less than M, by clustering the M filters and selecting a filter located at a center of each cluster.
4. The determination method according to claim 3 , wherein, in the merging, the M filters are clustered into L clusters by using k-means clustering.
5. The determination method according to claim 3 , wherein, in the merging, the M filters are clustered by using Affinity propagation clustering.
6. The determination method according to claim 1 , wherein the transformation includes a rotational transformation using a randomly determined angle, and
the splitting adds the at least one filter by performing the rotational transformation on the at least one of the N filters.
7. The determination method according to claim 1 , wherein the transformation includes an addition of Gaussian noise with randomly determined standard deviation, and
the splitting adds the at least one filter by performing the addition of Gaussian noise on the at least one of the N filters.
8. The determination method according claim 1 , wherein the transformation includes a contrast transformation for obtaining a randomly determined contrast ratio, and
the splitting adds the at least one filter by performing the contrast transformation on the at least one of the N filters.
9. The determination method according to claim 1 , wherein the transformation includes a scale transformation for obtaining a randomly determined scale, and
the splitting adds the at least one filter by performing the scale transformation on the at least one of the N filters.
10. A determination method for determining a structure of a convolutional neural network, the determination method comprising:
acquiring M filters having weights trained using a training image group as initial values, where M is a natural number greater than or equal to 2; and
merging the M filters into L filters, where L is a natural number greater than or equal to 1 and is less than M, by clustering the M filters and selecting a filter located at a center of each cluster.
11. A non-transitory computer-readable recording medium storing a program, the program causing a computer to execute a process that includes:
acquiring N filters having weights trained using a training image group as initial values, where N is a natural number greater than or equal to 1; and
splitting at least one of the N filters to increase a number of filters from N to M and to acquire M filters, where M is a natural number greater than or equal to 2 and is greater than N,
wherein the splitting adds at least one filter by performing a transformation used in image processing fields on each of the at least one of the N filters.
12. The determination method according claim 10 , wherein the M filters are merged into L clusters, with the filter located at a center of each of the L clusters being selected to define the L filters, in order to prevent overlearning.
13. The determination method according claim 10 , wherein the M filters are merged into L clusters, and
the determination method further comprises:
calculating a center of mass for each of the L clusters, with the filter being closest to the center of mass for each of the L clusters being selected to define the L filters.