Scholarlink

<p>Micro-expressions (MEs) are brief and involuntary facial expressions, posing a challenge for automatic classification. Although Convolutional Neural Networks (CNNs) can help with this task when ample training data are available, the size of the ME datasets is often limited. Furthermore, existing CNN models are deep and require long training times. In this study, we addressed these two problems by proposing a new framework called AuLiTENet comprising of two novel components: Augmented Video with Apex (AuViA) and Lightweight and Time Efficient Network (LiTENet) for ME classification. AuViA is a data augmentation technique where each ME video sample generates multiple small video samples. Each of these samples includes the apex frame, and they are transformed into single-instance images that effectively convey ME-related information. Conversely, the LiTENet is a CNN that incorporates collateral kernel streams of different sizes to capture salient ME features and perform classification. The major advantage of the proposed AuLiTENet framework is that the fixed number of kernels in each stream makes the network lightweight in parameter learning, reduces training time, and improves performance. We investigated the efficiency of the proposed AuLiTENet on five standard ME datasets and achieved superior performance compared to 11 existing works with respect to accuracy, F1 score, precision, recall, elapsed training time, and number of learning parameters.</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

AuLiTENet: a new Lightweight and Time-efficient CNN architecture for Micro-expression classification leveraging Data Augmentation

  • 期刊论文

  • 发表日期：

    2025年7月8日

查看全文
  • Ankita Jain,
  • Dhananjoy Bhakta,
  • Prasenjit Dey

摘要

Micro-expressions (MEs) are brief and involuntary facial expressions, posing a challenge for automatic classification. Although Convolutional Neural Networks (CNNs) can help with this task when ample training data are available, the size of the ME datasets is often limited. Furthermore, existing CNN models are deep and require long training times. In this study, we addressed these two problems by proposing a new framework called AuLiTENet comprising of two novel components: Augmented Video with Apex (AuViA) and Lightweight and Time Efficient Network (LiTENet) for ME classification. AuViA is a data augmentation technique where each ME video sample generates multiple small video samples. Each of these samples includes the apex frame, and they are transformed into single-instance images that effectively convey ME-related information. Conversely, the LiTENet is a CNN that incorporates collateral kernel streams of different sizes to capture salient ME features and perform classification. The major advantage of the proposed AuLiTENet framework is that the fixed number of kernels in each stream makes the network lightweight in parameter learning, reduces training time, and improves performance. We investigated the efficiency of the proposed AuLiTENet on five standard ME datasets and achieved superior performance compared to 11 existing works with respect to accuracy, F1 score, precision, recall, elapsed training time, and number of learning parameters.