Preprocessing Scheme of input images for Efficient Classification with Deep Learning

Abstract

Recently, deep learning-based artificial intelligence technology has significantly improved object recognition ability and provides efficient classification and decision results in various fields. For example, by photographing a food using a smartphone, it is possible to easily check the information of the food. However, the recognition performance varies highly depending on the quality and size of the picture taken by the user. In this paper, we propose a method to convert food packaging paper recognition into input data suitable for network calculation by performing a pre-processing process rather than the conventional detection method. This provides a result of improving the recognition rate by emphasizing the region to be detected through the image. For performance verification, we compared recognition rates with existing methods after training with different amount of datasets.