Scholarlink

<p>Local binary pattern (LBP) is a commonly adopted method for texture description in image analysis applications. However, when used in deep convolutional neural networks (CNNs), standard LBP often fails to retain crucial feature uniformity and diversity during dimensionality reduction, which limits its effectiveness in complex image classification tasks. To address these issues, a novel LBP encoding method utilizing balanced Gray code (BGC) is introduced. In this method, similar LBP features are re-ordered as neighbors to maximally preserve feature integrity and diversity through the pooling process. This approach improves the compatibility of LBP with CNN architectures, enhancing both accuracy and robustness. An efficient algorithm is designed that integrates BGC transformation, LBP counting, and sum pooling for dimensionality reduction without sacrificing key feature characteristics. Several LBP-embedded network architectures based on ResNet have been developed and tested, demonstrating significant performance improvements in the proposed method. The experimental results reveal that our proposed method outperforms typical CNN models with a 15% accuracy improvement for classifying the STL-10 benchmark images. Additionally, our proposed method demonstrates greater robustness to illumination variations, highlighting the potential of this method for enhanced feature representation in deep learning.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Integrating local binary patterns and convolutional neural networks with a balanced gray encoded pooling mechanism

  • 期刊论文

  • 发表日期：

    2025年4月10日

查看全文
  • Cheng-Ling Lai,
  • Ling-Shen Tseng,
  • Chih-Hung Chang,
  • Yi-Han Chen,
  • Chih-Hung Wu

摘要

Local binary pattern (LBP) is a commonly adopted method for texture description in image analysis applications. However, when used in deep convolutional neural networks (CNNs), standard LBP often fails to retain crucial feature uniformity and diversity during dimensionality reduction, which limits its effectiveness in complex image classification tasks. To address these issues, a novel LBP encoding method utilizing balanced Gray code (BGC) is introduced. In this method, similar LBP features are re-ordered as neighbors to maximally preserve feature integrity and diversity through the pooling process. This approach improves the compatibility of LBP with CNN architectures, enhancing both accuracy and robustness. An efficient algorithm is designed that integrates BGC transformation, LBP counting, and sum pooling for dimensionality reduction without sacrificing key feature characteristics. Several LBP-embedded network architectures based on ResNet have been developed and tested, demonstrating significant performance improvements in the proposed method. The experimental results reveal that our proposed method outperforms typical CNN models with a 15% accuracy improvement for classifying the STL-10 benchmark images. Additionally, our proposed method demonstrates greater robustness to illumination variations, highlighting the potential of this method for enhanced feature representation in deep learning.