• Search Research Projects
  • Search Researchers
  • How to Use
  1. Back to previous page

Extracting Fusion Features of Deep Convolutional Neural Networks Trained on Different Large-scale Datasets

Research Project

Project/Area Number 17K20008
Research Category

Grant-in-Aid for Challenging Research (Exploratory)

Allocation TypeMulti-year Fund
Research Field Human informatics and related fields
Research InstitutionKyushu University

Principal Investigator

Tetsu Matsukawa  九州大学, システム情報科学研究院, 助教 (80747212)

Project Period (FY) 2017-06-30 – 2021-03-31
Project Status Completed (Fiscal Year 2020)
Budget Amount *help
¥6,240,000 (Direct Cost: ¥4,800,000、Indirect Cost: ¥1,440,000)
Fiscal Year 2019: ¥1,560,000 (Direct Cost: ¥1,200,000、Indirect Cost: ¥360,000)
Fiscal Year 2018: ¥1,300,000 (Direct Cost: ¥1,000,000、Indirect Cost: ¥300,000)
Fiscal Year 2017: ¥3,380,000 (Direct Cost: ¥2,600,000、Indirect Cost: ¥780,000)
Keywords畳込ネット / 学習済みモデル / 特徴転移 / 正規直交化 / 融合 / 双線形プーリング / 画像識別 / カメラ間人物照合 / CNN特徴転移 / 判別分析 / 次元削減 / 畳込ニューラルネットワーク / 画像認識 / 深層学習 / 畳み込みニューラルネットワーク
Outline of Final Research Achievements

We have developed feature extraction methods from multiple pre-trained CNNs for transferring the learned representation to various image recognition tasks. First, we have shown that the orthonormalization of the weights of a fully connected layer improves the performance when transferring them to various tasks. We have extracted the localized feature maps of the orthonormalized fully connect layer from two CNNs, and fused them with bilinear pooling. The proposed fusion features achieved comparable or better performance compared with the fusion features of the top convolutional layers with a smaller computational cost. We have also developed a method, which conducts discriminative pooling of the convolutional features with training data of a target task.

Academic Significance and Societal Importance of the Research Achievements

クラス代表ベクトルの正規直交化は, 対象タスクの学習データを利用せず学習済みのモデルのみから実行可能であるが, 全結合層の特徴をそのまま利用した場合よりも対象タスクでの認識性能が高く,最上位畳み込み層よりも出力される次元数が少ない特徴抽出を可能とする. これを特徴融合に利用することで,計算量が削減された高性能な融合特徴を抽出できる.また, 判別的特徴集積は,小サンプルデータにおいて高速に学習でき, CNNをファインチューニングするよりも有効であった. このように本研究成果は, 学習済みの巨大な深層学習のモデルを, 限られた計算資源で小サンプルデータ問題に適用する場合に有用である.

Report

(5 results)
  • 2020 Annual Research Report   Final Research Report ( PDF )
  • 2019 Research-status Report
  • 2018 Research-status Report
  • 2017 Research-status Report
  • Research Products

    (4 results)

All 2021 2020 2019

All Journal Article (1 results) (of which Peer Reviewed: 1 results) Presentation (3 results) (of which Int'l Joint Research: 2 results)

  • [Journal Article] Hierarchical Gaussian Descriptor with Application to Person Re-Identification2020

    • Author(s)
      Tetsu Matsukawa, Takahiro Okabe, Einoshin Suzuki, Yoichi Sato
    • Journal Title

      IEEE Transactions on Pattern Analysis and Machine Intelligence

      Volume: PP Issue: 9 Pages: 1-16

    • DOI

      10.1109/tpami.2019.2914686

    • Related Report
      2019 Research-status Report
    • Peer Reviewed
  • [Presentation] Convolutional Feature Transfer via Camera-specific Discriminative Pooling for Person Re-Identification2021

    • Author(s)
      Tetsu Matsukawa, Einoshin Suzuki
    • Organizer
      25th International Conference on Pattern Recognition (ICPR2020)
    • Related Report
      2020 Annual Research Report
    • Int'l Joint Research
  • [Presentation] 正規直交クラス代表ベクトルとの内積表現によるConvNet特徴転移2021

    • Author(s)
      松川徹, 鈴木英之進
    • Organizer
      第24回画像の認識・理解シンポジウム (MIRU2021)
    • Related Report
      2020 Annual Research Report
  • [Presentation] Kernelized Cross-View Quadratic Discriminant Analysis for Person Re-Identification2019

    • Author(s)
      Tetsu Matsukawa, Einoshin Suzuki
    • Organizer
      Sixteenth International Conference on Machine Vision Applications (MVA 2019)
    • Related Report
      2019 Research-status Report
    • Int'l Joint Research

URL: 

Published: 2017-07-21   Modified: 2022-01-27  

Information User Guide FAQ News Terms of Use Attribution of KAKENHI

Powered by NII kakenhi