[3N4-J-10] Vision, speech: applications to industries

[3N4-J-10-02] Prediction of Favorability Rating on Beer-Can Package Designs Using Convolution Neural Network and Visualization by Class Activation Mapping.

〇Hiroyuki Shinohara¹, Tatsuji Ishiguro¹, Shunsuke Nakamura², Toshihiko Yamasaki² (1. Kirin Company, Limited, 2. The University of Tokyo)

Keywords:Deep Learning, Package Design

A quantitative survey of favorability rating by multiple panels is the main method to decide package designs. However, it is expensive and there is a limitation in the number of designs that can be subjected to a single survey. Therefore, this study aims at predicting the panel evaluation from the past survey results by a convolution neural network and visualize the important features by Grad-CAM. As a result, it has been made possible to give prescreening test to package design and suggestions of important features to the designers.