Machine learning method, machine learning device, and machine learning program

Abstract

A full-size training image is reduced by an image reduction unit () and input to an FCN (Fully Convolutional Neural Network) computation unit (), and the FCN computation unit () performs calculation under a set filter coefficient and outputs a reduced label image. The reduced label image is enlarged to a full size by an image enlargement unit (), the error calculation unit () calculates an error between the enlarged label image and a full-size ground truth based on a loss function, and the parameter update unit () updates a filter coefficient depending on the error. By repeating learning under the control of the learning control unit, it is possible to generate a learning model for performing optimal segmentation including an error generated at the time of image enlargement. Further, by including the image enlargement processing in the learning model, a full-size label image can be output, and the accuracy evaluation of the model can also be performed with high accuracy.