@techreport{oai:ipsj.ixsq.nii.ac.jp:00241567, author = {林, 正悟 and 山口, 智浩 and 安本, 慶一 and 松井, 智一 and Shogo, Hayashi and Tomohiro, Yamaguti and Keiithi, Yasumoto and Tomokazu, Matsui}, issue = {6}, month = {Dec}, note = {近年,宅内行動認識にはカメラや音声センサが活用されているが,カメラはプライバシ侵害の懸念があり,従来のマスキング処理では行動認識精度が低下する問題がある.本研究では,DeepPrivacy2 を用いて人物を架空の存在に置き換える手法に加え,背景も匿名化することで,画像全体のプライバシを保護しながら,画像内の人物の行動的特徴が加工前と比べてなるべく変化がないような画像生成システムを提案し,作成する.提案手法では撮影した画像を人物領域と背景領域に分割し,それぞれ別々の画像生成 AI を用いて匿名化を行い,匿名化した領域同士を元の位置に合成した画像を出力する.また,システム作成後に被験者実験を行い,提案手法がプライバシ保護に優れているかどうかを評価する.また行動認識用ツールを用いて加工前と加工後の画像をそれぞれ分析し,提案手法の匿名加工の定量的評価を行う., In recent years, in-home activity recognition has utilized cameras and audio sensors; however, cameras raise privacy concerns, and traditional masking techniques can degrade recognition accuracy. This study proposes and develops an image generation system that, by employing DeepPrivacy2 to replace individuals with fictitious entities and anonymizing the background, protects the privacy of the entire image while preserving the behavioral features of individuals as much as possible compared to the original image. The proposed method involves segmenting captured images into person and background regions, anonymizing each using separate image generation AIs, and then recombining the anonymized regions into their original positions to produce the final image. After system development, we will conduct experiments with participants to evaluate the effectiveness of the proposed method in privacy protection. Additionally, we will perform quantitative assessments of the anonymization by analyzing both the original and processed images using activity recognition tools.}, title = {人物領域と背景領域双方のプライバシ保護のための画像生成システムの作成}, year = {2024} }