Scholarlink

<p>Facial emotion recognition (FER) is a prevalent area of research in the field of computer vision, yet detecting a person's emotions in a complex environment remains challenging due to substantial intra-class variances. In current image classification approaches, much research has provided an endwise structure to recognize facial emotions using deep learning algorithms. Consequently, an approach for reducing the computational burden by decreasing the number of feature dimensions while boosting FER efficiency is proposed in this paper<i>.</i> MACS<sup>2</sup>_Net is a multipath hybrid deep learning framework constructed of two pathways. Path-1 (DA_CNN) is an antialiasing-based CNN model that focuses on capturing local patterns, in which antialiasing is incorporated to eliminate intermediate artifacts arising from downsampling steps. Path-2 (DS_SN) is a StochasticDepth-based SqueezeNet model that relies on capturing more global context. The StochasticDepth layer is chosen for its capacity to effectively cope with sparse and noisy inputs, limiting overfitting while offering effective computing. When merged, they can provide more comprehensive representations to assist with classifying emotions. Davis’s Library (Dlib) is utilized at first to extract facial landmarks with the goal of reducing redundancy in facial data. Following that, these landmarks will be simultaneously delivered as input for both Path-1 and Path-2. Each of these models have then been evaluated on lab-controlled datasets and wild FER dataset to detect eight distinct emotions from image data. The observations from the proposed models will then be contrasted against one another. Furthermore, several performance measures involving accuracy, precision, specificity, sensitivity, Jaccard coefficient, training time, and the overall number of parameters have been analyzed to assess the reliability and efficiency of the proposed methods.&#xa0;The multipath MACS<sup>2</sup>_Net framework outperforms all other classification models both in terms of reliability and precision while reducing the depth of the input data, alleviating training time, with an accuracy of 99.3%, 83.4%, and 98.6% for the Extended Cohn-Kanade (CK +), Real-world Affective Face (RafD), and Japanese female facial expressions (Jaffee) data, respectively.</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

MACS2_Net: a stochastic depth with antialiasing on an optimized multipath convolutional-SqueezeNet framework for facial emotion recognition

  • 期刊论文

  • 发表日期：

    2025年8月6日

查看全文
  • Reham A. Elsheikh,
  • M. A. Mohamed,
  • Ahmed Mohamed Abou-Taleb,
  • Mohamed Maher Ata

摘要

Facial emotion recognition (FER) is a prevalent area of research in the field of computer vision, yet detecting a person's emotions in a complex environment remains challenging due to substantial intra-class variances. In current image classification approaches, much research has provided an endwise structure to recognize facial emotions using deep learning algorithms. Consequently, an approach for reducing the computational burden by decreasing the number of feature dimensions while boosting FER efficiency is proposed in this paper. MACS2_Net is a multipath hybrid deep learning framework constructed of two pathways. Path-1 (DA_CNN) is an antialiasing-based CNN model that focuses on capturing local patterns, in which antialiasing is incorporated to eliminate intermediate artifacts arising from downsampling steps. Path-2 (DS_SN) is a StochasticDepth-based SqueezeNet model that relies on capturing more global context. The StochasticDepth layer is chosen for its capacity to effectively cope with sparse and noisy inputs, limiting overfitting while offering effective computing. When merged, they can provide more comprehensive representations to assist with classifying emotions. Davis’s Library (Dlib) is utilized at first to extract facial landmarks with the goal of reducing redundancy in facial data. Following that, these landmarks will be simultaneously delivered as input for both Path-1 and Path-2. Each of these models have then been evaluated on lab-controlled datasets and wild FER dataset to detect eight distinct emotions from image data. The observations from the proposed models will then be contrasted against one another. Furthermore, several performance measures involving accuracy, precision, specificity, sensitivity, Jaccard coefficient, training time, and the overall number of parameters have been analyzed to assess the reliability and efficiency of the proposed methods. The multipath MACS2_Net framework outperforms all other classification models both in terms of reliability and precision while reducing the depth of the input data, alleviating training time, with an accuracy of 99.3%, 83.4%, and 98.6% for the Extended Cohn-Kanade (CK +), Real-world Affective Face (RafD), and Japanese female facial expressions (Jaffee) data, respectively.