画像ビッグデータと大規模被験者アンケートに基づく住宅地における街路印象評価推定モデル

抄録

This paper constructed a general-purpose deep learning model, which can evaluate the impression of arbitrary streets in residential areas from various aspects. First, we conducted a Web-based street impression evaluation questionnaire (22 items) with 14,900 participants using 1,000 streetscape images from Google Street View. Next, we trained a deep learning-based ranking learning model using the images and questionnaire results. The F-value of the trained model was about 77% on average. Furthermore, we analyzed the correlations among evaluation items and discussed the spatial distribution of impressions in Tokyo 23 wards based on the scores estimated by the trained model.