Method for improving the efficiency of preparing open image data using deep learning

Open data release and utilization is positioned as a first step in advancing regional / municipal digital transformation (DX). In the tourism sector, publishing open data can enhance promotional activities through public-private partnerships. Among various data types, open image data holds substantial potential due to its ability to instantly capture public interest with attractive visuals. However, compared to tabular data, image-based open data lacks established templates and precedents, making the publication process unclear and underutilized by local government officials. To address this issue, this study proposed and developed a support process for preparing open image data, leveraging deep learning techniques. The process consists of three key stages: candidate image extraction, metadata generation, and the development of an application for manual verification. In the first stage, we extracted candidate images by scoring a dataset using a model developed in previous research to assess the attractiveness of images. Images exceeding a predetermined threshold were selected. Next, cosine similarity between vectorized images was calculated to group and eliminate redundant images, retaining only one representative per group. We also implemented face detection to exclude images containing identifiable individuals, addressing privacy concerns. Additionally, folders containing images with copyright restrictions were specified and excluded from the open data set. These steps resulted in the selection of candidate images from a vast number of publicity images. In the second stage, generative AI was used to assign titles and tags to the selected images for metadata generation. In the third stage, an application was developed to allow local government officials to easily review, edit, and manage the images and their metadata via a CSV file. Local government staff highly evaluated the method for its effectiveness and efficiency, noting that prior consultations on thresholds and metadata rules contributed to an acceptable selection process.