Architecture has been disseminated in society through photography, as much as actual visits. In recent years, art museums have been positively encouraging visitors to take pictures, and the pictures taken by visitors are accumulated on social media as Big Data, showing citizen's architectural experience. This research aims at the following two points for big data formed by the accumulated posted pictures on "Instagram", a social media specialized for picture posting. (1)To visualize big data of images on museums. (2)To discover the way how to analyze big data that will help art museum plans. As a survey method, a picture analysis using deep learning was performed. First, a feature amount extraction of the pictures is performed by CNN, and then we visualized it using t-SNE to show the result of CNN in two dimensions and "image discrimination program" which mechanically classifies pictures for learned elements.
First, as the target of the survey, 595,587 pictures related to 406 museums were extracted from Instagram by hashtag search. In order to grasp the rough trend of the pictures, t-SNE was conducted. As a result, we have obtained five elements of "building" "art" "text" "nature" "food". (Fig. 2)
Second, we made "image discrimination program" learn about 5 elements, and the program identified the 595,587 pictures. The classification result of the pictures was "building" 13.7%, "art" 56.0%, "text" 20.0%, "nature" 6.2%, "food" 4.1%. In the comparison between the museum's outline and the 5 elements ratio, the numbers of picture posts and the numbers of visitors which often used for museum evaluation are not related (Fig. 4). Moreover, there are many "building" picture posts at the old museums of the opening year, "art" and "text" picture posts at the new museums (Fig. 6). From typing based on the ratio of 5 elements in each museum, we were able to grasp the characteristics of each building by obtaining 16 categories and 6 groups of types (Fig. 8, Table1). From the ratio of 5 elements on two museums holding the same exhibition which allows visitors to take pictures, we clarified the difference of the influence of the camera allowed exhibition(Fig. 9).
Third, we conducted t - SNE on the building pictures of each museum to visualize the detailed subjects. As a result, it was found that the images of a museum building on social media are made up of relationships between other elements (Fig. 10, 11). Moreover, as the ratio of building pictures at a museum increases, building pictures image is biased towards limited subject (Fig. 12).
These findings obtained by this research are not to show subjectivity reflected in each picture but to show a part of the characteristics and impressions of the museums from the accumulated image data by the visitors. Fumio Nanjo, Mori Art Museum Director said about the museum buildings, "It is not a model as a general theory, but a special solution that matches its context is needed". In other words, these data are not to indicate the majority to get the right answer. These data are more like the air showing how people get interested in and what they do there. By visualizing it, we suggest to building planners and designers that they can think about building architecture with feeling more opinions. We will position this research as an indication of options for each art museum to acquire its unique special solutions.
