Effective use of evolutionary adversarial learning and temporal feature to improve deep learning performatnce

2022 Fiscal Year Final Research Report

• Project/Area Number: 19K11515
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 59020:Sports sciences-related
• Research Institution: Iwate University
• Principal Investigator: Akashi Takuya 岩手大学, 理工学部, 准教授 (50403655)
• Project Period (FY): 2019-04-01 – 2023-03-31
• Keywords: スポーツ科学 / ビッグデータ / 進化計算 / 深層学習 / コンピュータビジョン

Outline of Final Research Achievements

The main objective of this paper is to establish a novel approach that integrates spatial and temporal features in order to achieve a behavior and emotion estimation system that closely resembles human cognitive abilities.
In the first year, we proposed a feature matching method for 3D point clouds, and a multimodal learning approach using audio and image data and published it as a journal paper. In the second year, we focused on researching new matching schemes for feature descriptors and proposed a novel method using deep learning and transfer learning. This research was also published as a journal paper. In the third year, we published a journal paper on the detection of partially occluded human heads. In the final year, we are further exploring the fusion of spatial and temporal feature descriptors and present at international conference. We are also in the process of writing a journal paper on this topic.

Free Research Field

コンピュータビジョン

Academic Significance and Societal Importance of the Research Achievements

コンピュータビジョン分野においては，国内外の物体の認識手法として，深層学習が脚光を浴びているが，空間的な情報を用いるものがほとんどで，人体の関節など重要な特徴点の移動量の時間的変化を扱う研究は申請者の知る限り存在しない．さらに，空間的特徴量と時間的特徴量を組み合わせる手法も類を見ない．
本研究では，深層学習等の人工知能における空間的特徴と時間的特徴を融合に関して研究・開発した．このような手法が確立されれば，より人間の認知能力に近い行動・情動推定システム等が実現され，コンピュータサイエンス分野やニューロサイエンス分野におけるブレイクスルーにつながると考えられる．