A Study on The Flower Species Recognition Method Using Meta Infomation for Class

Abstract

In recent years, many methods using deep learning have been proposed and have shown good results in image recognition tasks. Furthermore, improvement of recognition accuracy has been reported by using information attached to each data (metadata) for learning, such as multitask learning. However, the task of annotating metadata to each data is high cost, and the cost increases in proportion to the number of data. In this study, we improve the image recognition accuracy by using the information given to the prediction class as metadata. In addition, we propose the loss function to avoid inconsistency of information with image and metadata caused by using information to classes as metadata. In the experiment, we verified the effectiveness of our method using flower images posted to our service and the color information given for each flower classes, and confirmed 1.49% improvement in accuracy over the comparison method.