A study on acquisition of intermediate representations using context and top-down information for image recognition

2020 Fiscal Year Final Research Report

• Project/Area Number: 16K00239
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Perceptual information processing
• Research Institution: Hiroshima University
• Principal Investigator: Kurita Takio 広島大学, 先進理工系科学研究科(工), 教授 (10356941)
• Co-Investigator(Kenkyū-buntansha): 日高 章理 東京電機大学, 理工学部, 准教授 (70553519)
• Project Period (FY): 2016-04-01 – 2021-03-31
• Keywords: 画像認識 / 機械学習 / 人工知能 / 深層学習

Outline of Final Research Achievements

As methods for learning models with high generalization performance, we proposed a mix-up learning method using feature vectors in the middle layer, and a method for visualizing trained networks and feature vectors using canonical correlation analysis. We also proposed a method for extracting vascular regions from retinal images that uses Euler's polyhedron theorem to estimate the number of connected components and guides learning to minimize it. Furthermore, in order to incorporate contextual information into the learning process, we proposed a method for constructing feature vectors for similar image retrieval by learning from the features of the network learned for object detection using Triplet Loss.

Free Research Field

情報科学

Academic Significance and Societal Importance of the Research Achievements

深層学習を用いた画像認識の性能や適用領域の拡大を目指して，（１）画像認識に有効な特徴の学習法に関する研究，（２）トップダウン情報を利用した中間表現の獲得に関する研究，（３）文脈情報を利用した中間表現の獲得に関する研究を行い，トップダウン情報や文脈情報を学習に取り込むことの重要性を再確認した．