Investigation on the Features Extracted by CNN Kernels from Images

DOI

抄録

CNN (Convolutional Neural Network) is a type of neural network that shows excellent performance in tasks such as image recognition and object detection. CNNs are generally said to extract local features using ker- nels. In this paper, we first create an artificial image dataset and confirm its features. We then use this dataset to train a simple structured CNN and investigate what kind of fea- tures it actually extracts. By changing the kernel size of the CNN and performing image classification, we find that changes in classification accuracy vary by class. This sug- gesta that CNNs capture local features rather than over- all features of images. Additionally, we have found that padding may affect the features extracted.

収録刊行物

  • IEICE Proceeding Series

    IEICE Proceeding Series 76 513-516, 2023-09-21

    The Institute of Electronics, Information and Communication Engineers

キーワード

詳細情報 詳細情報について

  • CRID
    1390016372307961472
  • DOI
    10.34385/proc.76.c3l-12
  • ISSN
    21885079
  • 本文言語コード
    en
  • データソース種別
    • JaLC
  • 抄録ライセンスフラグ
    使用不可

問題の指摘

ページトップへ