マルチパッチABNに基づく写真の審美的品質推定

カテゴリ: 部門大会

論文No: OS7-1-2

グループ名: 【C】2023年電気学会電子・情報・システム部門大会

発行日: 2023/08/23

タイトル(英語): Aesthetic Quality Estimation of Photographs based on Attention Branch Network and Multi-patch

著者名: 坂口 太一（岡山県立大学）,滝本 裕則（岡山県立大学）,金川 明弘（岡山県立大学）

著者名(英語): Daichi Sakaguchi (Graduate school of Okayama Prefectural University),Hironori Takimoto (Okayama Prefectural University),Akihiro Kanagawa (Okayama Prefectural University)

キーワード: 審美的品質推定|Attention branch network|深層学習深層学習|Aesthetic quality estimation|Attention branch network|Deep learning

要約(日本語): Aesthetics assessment of photographs has attracted attention because of its potential use in various applications, including image retrieval, cropping, and photo enhancement. However, it is challenging due to its fuzzy definition and highly subjective nature. This paper proposes an aesthetic quality estimation with a visual explanation function to improve the estimation accuracy. First, an estimation model based on the attentional branch network is used as a baseline to estimate the aesthetic quality of photographs. Second, to improve the estimation accuracy, features extracted from multiple patches are used for training. Finally, from the experimental results using the benchmark dataset AVA, it is confirmed that the proposed method improves the accuracy of aesthetic quality estimation.