Modeling Kansei Index for Images and Impression Estimation Using Fine Tuning

Visual impression estimation system considering attribute information

“…Machine learning-based indexing methods using review data can automatically generate a large amount of data, thus reducing the human and time load, but have the problems that a large amount of domain review data cannot always be obtained and that the review data does not always cover the impressions received by consumers. In a study by the authors [7] that attempted to solve these problems, the evaluation data covering impressions were collected to construct Kansei index while reducing human and temporal load, and Kansei index was estimated using a deep learning model. However, this study estimated the average impressions of respondents and did not address individual or attribute trends.…”

“…In this section, we first describe the dataset creation, followed by the model training. We create the dataset in the same way as in previous research [7] . The dataset is created by first defining the Kansei evaluation axis, which is the evaluation factor of the domain, and then calculating the evaluation values for each attribute of the respondent's attributes.…”

“…To create the dataset, 1,000 lunchbox images from a previous study [7] , evaluation data for 32 words and factor analysis results for the evaluation data were used. We also used gender as an attribute.…”

“…Therefore, this study proposes a method to infer the impressions received from images for each respondent's attributes. The target domain is set as Japanese lunchbox and Kansei evaluation axis is used as the evaluation factor for lunchbox from previous research [7] . Using the collected questionnaires and factor analysis results from the previous study, the evaluation values for each evaluation axis are calculated for each respondent's attribute and a dataset is created.…”