Scholarlink

<p>Facial expression recognition (FER) is a significantly important task in the extended vision community, having extensive applications in human-computer interaction, and more so in generative vision models. Moreover, its application majorly lies in lightweight devices such as smartphones, CCTV cameras, robots etc. and thus requires models that can deployed on low compute. The challenge, hence, lies in learning an effective feature representation that is both robust and computationally efficient. In this work, we propose a two-stage pipeline for FER, that harnesses the power of deep transfer learning and statistical feature selection to achieve the aforementioned goals. Specifically, in the first stage we adopt a pre-trained EfficientNet model and fine-tune it on our target dataset, followed by extracting high-dimensional features using the frozen backbone. Once extracted, we leverage information gain, a statistical measure based on entropy differences, to quantify the “usefulness” of each of the features and rank them. Taking the top-<i>k</i> subset from the ranked features, we train a k-nearest neighbour classifier to perform the final classification. Upon evaluation, our method proves to be highly competitive, outperforming several existing state-of-the-art works by significant margins on three commonly used FER datasets.</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

A Lightweight Deep Feature Selection Contour for Emotion Recognition from Human Faces

  • 期刊论文

  • 发表日期：

    2025年9月11日

查看全文
  • Sabyasachi Tribedi,
  • Ranjit Kumar Barai

摘要

Facial expression recognition (FER) is a significantly important task in the extended vision community, having extensive applications in human-computer interaction, and more so in generative vision models. Moreover, its application majorly lies in lightweight devices such as smartphones, CCTV cameras, robots etc. and thus requires models that can deployed on low compute. The challenge, hence, lies in learning an effective feature representation that is both robust and computationally efficient. In this work, we propose a two-stage pipeline for FER, that harnesses the power of deep transfer learning and statistical feature selection to achieve the aforementioned goals. Specifically, in the first stage we adopt a pre-trained EfficientNet model and fine-tune it on our target dataset, followed by extracting high-dimensional features using the frozen backbone. Once extracted, we leverage information gain, a statistical measure based on entropy differences, to quantify the “usefulness” of each of the features and rank them. Taking the top-k subset from the ranked features, we train a k-nearest neighbour classifier to perform the final classification. Upon evaluation, our method proves to be highly competitive, outperforming several existing state-of-the-art works by significant margins on three commonly used FER datasets.