Study on Image Representation Learning and Understanding based on Human's Perception Principle and Deep Statistical Analysis

2017 Fiscal Year Final Research Report

• Project/Area Number: 15K00253
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Perceptual information processing
• Research Institution: Yamaguchi University (2016-2017); Ritsumeikan University (2015)
• Principal Investigator: Han Xian-HUa 山口大学, 大学院創成科学研究科, 准教授 (60469195)
• Co-Investigator(Kenkyū-buntansha): 陳 延偉 立命館大学, 情報理工学部, 教授 (60236841)
• Project Period (FY): 2015-04-01 – 2018-03-31
• Keywords: 画像認識 / 機械学習 / コンピュータビジョン / パターン認識 / 画像処理

Outline of Final Research Achievements

This study aimed at learning compact and inherent image representation for high-level vision probelms, and developed advanced image recognition and understanding methods. Our main achievements are three-fold: 1) Based on human’s perception principle, we transformed the raw-image domain into differential excitation domain and proposed to use the micro-texton as local descriptors for retaining all information, which would be distinguishable even for the subtle difference in image structures. 2) We proposed a novel middle-level image representation learning framework via exploring the deviation statistics of local descriptor set on the learned GMM model; 3) We stacked several layers of the middle level representation extraction framework, and proposed multiple-layer fisher network architecture for high-level feature learning. We applied our proposed image representation learning strategy for several image recognition applications, and proved much better performances can be achieved.

Free Research Field

知覚情報処理