Prior learning by deep epigraphical network and its application to rare imaging

2023 Fiscal Year Final Research Report

• Project/Area Number: 21K04045
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 21020:Communication and network engineering-related
• Research Institution: Kogakuin University
• Principal Investigator: Kyochi Seisuke 工学院大学, 情報学部(情報工学部), 准教授 (70634616)
• Project Period (FY): 2021-04-01 – 2024-03-31
• Keywords: 画像復元 / エピグラフ緩和 / 深層ニューラルネットワーク / 凸最適化

Outline of Final Research Achievements

This project established a image recovery algorithm, which estimates true images from degraded images with noise and other artifacts produced during the measurement process, called "Deep Epigraphical Networks" that is more effective than deep neural networks when training dataset is not available (such as medical and industrial images). As an image restoration method that does not require big data for training, estimation through the minimization of regularization functions designed based on prior information is widely utilized. In this research, we demonstrated the practicality of our approach by solving the design method for deep composite regularization that models the complex prior information of target image data and its minimization algorithm using the principal investigator's unique convex optimization technique, "Epigraph Deformation."

Free Research Field

信号処理・機械学習

Academic Significance and Societal Importance of the Research Achievements

本研究の成果である「深層エピグラフネットワーク」は学習用ビッグデータが必要ないため，これまで深層ニューラルネットワークの適用の難しかった，医療分野や産業分野の画像データ（CT・PET・ハイパースペクトル画像等）への応用が期待できる．本技術により，ノイズや欠損，ボケ，ブレなどの劣化が生じやすい画像の品質向上が実現できるため，医用現場では診断の精度向上や患者の治療効果の向上が期待され，産業方面では，例えば製造業などにおける品質管理の自動化・精度向上にも寄与できる．学術的にもインパクトは大きく，計測工学や生命科学など様々な学術領域の発展に寄与すると期待できる.