Scholarlink

<p>Creating efficient and lightweight models is crucial for the deployment of automatic facial expression recognition systems in resource-constrained environments like mobile devices and embedded systems. Faces occluded by masks pose an additional challenge to traditional facial expression analysis systems due to missing or hidden information associated with the obscured lower part of the face. Novel deep learning architectures need to be devised that pay more attention to the visible part of the face during training. To address these concerns, this study proposes an enhanced MobileNetV2 architecture incorporating a stack of three attention blocks and three squeeze-and-excitation (SE) blocks at the network output. This method preserves the lightweight architecture of MobileNetV2 while augmenting its feature extraction efficacy, offering a resource-efficient option for implementation in mobile or edge devices and low-power applications. In resource-constrained situations, the model excels due to its low computing cost and excellent accuracy, as evidenced by experimental findings on the benchmark masked facial emotion dataset Masked Extended Cohn-Kanade (MCK +) for fear, contempt, disgust, anger, happiness, sadness, and surprise emotions. The attention-augmented and SE-enhanced MobileNetV2 model achieves an accuracy of 97.94 ± 4.59%, AUC of 0.9997, F1-Score of 0.9796, Cohen’s Kappa of 0.9748, IBA of 0.9584, Precision of 0.98, Recall of 0.98 and MCC of 0.9750, surpassing state-of-the-art deep learning models like InceptionV3, ResNet50, EfficientNetB0, DenseNet121, Xception, NasNet, CNN, as well as other approaches on face emotion recognition. Extensive experiments and ablation studies on several plausible architectural expansions and augmentations of the MobileNetV2 architecture establish the supremacy of the proposed approach.</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

Attention-augmented squeeze-and-excitation enhanced mobile network for occluded facial expression recognition in resource-constrained environments

  • 期刊论文

  • 发表日期：

    2025年6月5日

查看全文
  • Ashi Agarwal,
  • Seba Susan

摘要

Creating efficient and lightweight models is crucial for the deployment of automatic facial expression recognition systems in resource-constrained environments like mobile devices and embedded systems. Faces occluded by masks pose an additional challenge to traditional facial expression analysis systems due to missing or hidden information associated with the obscured lower part of the face. Novel deep learning architectures need to be devised that pay more attention to the visible part of the face during training. To address these concerns, this study proposes an enhanced MobileNetV2 architecture incorporating a stack of three attention blocks and three squeeze-and-excitation (SE) blocks at the network output. This method preserves the lightweight architecture of MobileNetV2 while augmenting its feature extraction efficacy, offering a resource-efficient option for implementation in mobile or edge devices and low-power applications. In resource-constrained situations, the model excels due to its low computing cost and excellent accuracy, as evidenced by experimental findings on the benchmark masked facial emotion dataset Masked Extended Cohn-Kanade (MCK +) for fear, contempt, disgust, anger, happiness, sadness, and surprise emotions. The attention-augmented and SE-enhanced MobileNetV2 model achieves an accuracy of 97.94 ± 4.59%, AUC of 0.9997, F1-Score of 0.9796, Cohen’s Kappa of 0.9748, IBA of 0.9584, Precision of 0.98, Recall of 0.98 and MCC of 0.9750, surpassing state-of-the-art deep learning models like InceptionV3, ResNet50, EfficientNetB0, DenseNet121, Xception, NasNet, CNN, as well as other approaches on face emotion recognition. Extensive experiments and ablation studies on several plausible architectural expansions and augmentations of the MobileNetV2 architecture establish the supremacy of the proposed approach.