Image recognition device applying transfer learning to an object recognition model

1 . An image recognition device comprising:

an object recognition unit that recognizes an object in an input image by using an object recognition model;

a recognition precision determination unit that determines a precision of recognition of the object in the input image;

a supervised image conversion unit that converts the input image for which the precision of recognition of the object is lower than a predetermined threshold value into a supervised image by labeling the input image based on a feature amount of the input image; and

a transfer learning unit that applies transfer learning to the object recognition model by using the supervised image as training data to update the object recognition model; and

a database that maps a feature amount in an intermediate layer toward an end of a neural network, which is an intermediate output provided when the input image is input to the object recognition model, to a label of the input image,

wherein the supervised image conversion unit refers to the database to acquire a label corresponding to a feature amount most similar to the feature amount of the input image and automatically converts the input image to the supervised image by attaching the acquired label to the input image for which the precision of recognition of the object is lower than the predetermined threshold value, and outputs the label of a recognition result of the input image for which the precision of recognition of the object is equal to or higher than the predetermined threshold value,

wherein the transfer learning unit adds the supervised image to a data set used in prior learning of the object recognition model to configure a new data set and applies transfer learning to the object recognition model by using the new data set as training data, and

wherein when a predetermined number of supervised images having the same label are collected, the transfer learning unit adds the supervised images to the data set used in the prior learning of the object recognition model to create the new data set.

2 . An image recognition method comprising:

recognizing an object in an input image by using an object recognition model;

determining a precision of recognition of the object in the input image;

converting the input image for which the precision of recognition of the object is lower than a predetermined threshold value into a supervised image by labeling the input image based on a feature amount of the input image;

applying transfer learning to the object recognition model by using the supervised image as training data to update the object recognition model; and

referring to a database that maps a feature amount in an intermediate layer toward an end of a neural network, which is an intermediate output provided when the input image is input to the object recognition model, to a label of the input image,

wherein the converting refers to the database to acquire a label corresponding to a feature amount most similar to the feature amount of the input image and automatically converts the input image to the supervised image by attaching the acquired label to the input image for which the precision of recognition of the object is lower than the predetermined threshold value, and outputs the label of a recognition result of the input image for which the precision of recognition of the object is equal to or higher than the predetermined threshold value,

wherein the applying transfer learning adds the supervised image to a data set used in prior learning of the object recognition model to configure a new data set and applies transfer learning to the object recognition model by using the new data set as training data, and

wherein when a predetermined number of supervised images having the same label are collected, the applying transfer learning adds the supervised images to the data set used in the prior learning of the object recognition model to create the new data set.

3 . A non-transitory computer-readable recording medium having embodied thereon an image recognition program that causes a computer to perform:

recognizing an object in an input image by using an object recognition model;

determining a precision of recognition of the object in the input image;

converting the input image for which the precision of recognition of the object is lower than a predetermined threshold value into a supervised image by labeling the input image based on a feature amount of the input image;

applying transfer learning to the object recognition model by using the supervised image as training data to update the object recognition model; and

referring to a database that maps a feature amount in an intermediate layer toward an end of a neural network, which is an intermediate output provided when the input image is input to the object recognition model, to a label of the input image,

wherein the converting refers to the database to acquire a label corresponding to a feature amount most similar to the feature amount of the input image and automatically converts the input image to the supervised image by attaching the acquired label to the input image for which the precision of recognition of the object is lower than the predetermined threshold value and outputs the label of a recognition result of the input image for which the precision of recognition of the object is equal to or higher than the predetermined threshold value,

wherein the applying transfer learning adds the supervised image to a data set used in prior learning of the object recognition model to configure a new data set and applies transfer learning to the object recognition model by using the new data set as training data, and

wherein when a predetermined number of supervised images having the same label are collected, the applying transfer learning adds the supervised images to the data set used in the prior learning of the object recognition model to create the new data set.