Scholarlink

<p>Recent advancements in human–computer interaction have rendered affective computing a prominent study area, focused on creating computational systems capable of comprehending and responding to human facial expressions. Despite promising accuracy results for Facial Expression Recognition (FER) in controlled environments, analyzing facial expressions at a distance with low-quality images leads to a significant drop in recognition accuracy. The state-of-the-art methods focused on FER in the presence of partial occlusion, spontaneous expressions, micro-expressions, and variations in illumination, pose, &amp; image resolution. The proposed work addresses the FER problem in the presence of mild and uniform blur, a critical issue unexplored in the literature.&#xa0;Blur invariants provide an alternate way to describe and recognize facial expressions in blurred images without the need for restoration or data augmentation. The main objective of the proposed method is to&#xa0;derive the blur invariant feature space from the prominent facial regions to mitigate the challenges introduced by the mild and uniform blur on expression recognition. The&#xa0;prominent&#xa0;contributions of the proposed work&#xa0;are,&#xa0;the Local Phase Quantization with Fourier Transform (LPQ-FT) feature to derive the discriminative features that offer a higher degree of robustness in the presence of moderate blur; the multi-stage Stacked Deep Convolutional Autoencoder employed to enhance the discriminative power of the feature distribution by deriving the abstract feature space from the blur insensitive feature space; Multiclass Support Vector Machine classifier to predict the expression class of the subject.&#xa0;The proposed work&#xa0;is implemented&#xa0;on three benchmark datasets&#xa0;and&#xa0;a five-fold cross-validation method&#xa0;is applied.&#xa0;The hybrid use of LPQ-FT for feature extraction and deep autoencoders for dimensionality reduction introduces a novel way to deal with the high-dimensionality of texture features while ensuring robustness to low intensity blur. The comparative analysis of misclassification rates between the proposed blur invariant feature and widely deployed state-of-the-art feature descriptors demonstrates the substantial performance gains of the proposed model. The generalization capability of the proposed feature descriptor robustness to blur indicates that it will be a better choice for real-time implementation of FER in&#xa0;the presence of&#xa0;mild and uniform blur.</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

Robustness of facial expression recognition systems in the presence of mild and uniform blur

  • 期刊论文

  • 发表日期：

    2025年7月1日

查看全文
  • Naveen Kumar H N,
  • Chandrashekar M Patil,
  • Guru Prasad M S,
  • Sharath K R,
  • Praveen Gujjar J

摘要

Recent advancements in human–computer interaction have rendered affective computing a prominent study area, focused on creating computational systems capable of comprehending and responding to human facial expressions. Despite promising accuracy results for Facial Expression Recognition (FER) in controlled environments, analyzing facial expressions at a distance with low-quality images leads to a significant drop in recognition accuracy. The state-of-the-art methods focused on FER in the presence of partial occlusion, spontaneous expressions, micro-expressions, and variations in illumination, pose, & image resolution. The proposed work addresses the FER problem in the presence of mild and uniform blur, a critical issue unexplored in the literature. Blur invariants provide an alternate way to describe and recognize facial expressions in blurred images without the need for restoration or data augmentation. The main objective of the proposed method is to derive the blur invariant feature space from the prominent facial regions to mitigate the challenges introduced by the mild and uniform blur on expression recognition. The prominent contributions of the proposed work are, the Local Phase Quantization with Fourier Transform (LPQ-FT) feature to derive the discriminative features that offer a higher degree of robustness in the presence of moderate blur; the multi-stage Stacked Deep Convolutional Autoencoder employed to enhance the discriminative power of the feature distribution by deriving the abstract feature space from the blur insensitive feature space; Multiclass Support Vector Machine classifier to predict the expression class of the subject. The proposed work is implemented on three benchmark datasets and a five-fold cross-validation method is applied. The hybrid use of LPQ-FT for feature extraction and deep autoencoders for dimensionality reduction introduces a novel way to deal with the high-dimensionality of texture features while ensuring robustness to low intensity blur. The comparative analysis of misclassification rates between the proposed blur invariant feature and widely deployed state-of-the-art feature descriptors demonstrates the substantial performance gains of the proposed model. The generalization capability of the proposed feature descriptor robustness to blur indicates that it will be a better choice for real-time implementation of FER in the presence of mild and uniform blur.