Convolutional sparse representation of l1 norm error criterion and its development for distributed video coding and deep learning

2023 Fiscal Year Final Research Report

• Project/Area Number: 20K11878
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 61010:Perceptual information processing-related
• Research Institution: Kurume National College of Technology
• Principal Investigator: Kuroki Yoshimitsu 久留米工業高等専門学校, 制御情報工学科, 教授 (60290847)
• Project Period (FY): 2020-04-01 – 2024-03-31
• Keywords: 畳み込み型スパース表現 / 分散圧縮符号化 / 深層学習 / 凸最適化

Outline of Final Research Achievements

The recent advancements in AI were triggered by convolutional neural networks surpassing existing methods in the image classification contest. This study focuses on convolutional sparse representations, which approximate images using a sum of convolutional filters and their coefficients. Here, "sparse" means that the number of filter coefficients contain many zeros as possible. If the approximation accuracy is the same, a higher sparsity is considered to better representation of image features. In this study, we examined the application of distributed compression coding and convolutional neural networks and have achieved results superior to conventional methods.

Free Research Field

知覚情報処理

Academic Significance and Societal Importance of the Research Achievements

本研究では畳み込みスパース表現のスパース性と近似精度の双方を向上させる方法として近似精度をL1ノルムと称する誤差ベクトルの絶対値和で評価する手法を提案した．また，計算負荷を低減し，大規模データで適用可能なコンセンサス方式を導出した．これらの成果は分散圧縮符号化および小規模な畳み込みニューラルネットワークにおける精度向上へとつながり，国際会議ICIIBMSにおけるStudent Best Paper Awardとして評価された．