機械学習によるRAW現像技術の開発

Abstract

RAW development is an operation manually performed by a person in order to finish a photograph according to his / her desire. At this time, many image adjustment parameters are set, and confirmation and adjustment are repeated. The work requires much time and effort, and it is very difficult to make many photos taken into the desired images. However, if deep learning is used, learning RAW images and photographs after RAW development may create photographs of the same level as manual RAW development. In this research, we succeeded in developing a machine learning model that performs RAW development using RAW data as input. We also aimed to improve the processing speed of learning using GPGPU. As a result of comparing processing in which RAW development is performed for 100 images with a single multi-core CPU using a parallel program and processing using a GPGPU, it is shown that the latter can be significantly faster.