Scholarlink

<p>Human behavior analysis significantly depends on facial expression recognition, where deep learning has enabled the development of visionary models that can surpass human-level performance. Explainable Artificial Intelligence (XAI) techniques are employed to validate the trustworthiness of a trained convolutional neural network by providing interpretable heatmaps, generated using recent techniques, including GradCAM, GradCAM++, LayerCAM, and ScoreCAM. These saliency heatmaps highlight the critical facial regions used by the classifiers, thus aligning the system’s behavior with human cognitive processes. Metrics such as average drop, confidence increase, and win percentage are utilized to assess the system’s reliability by analyzing these heatmaps, but they can not quantify the measure of trustworthiness. This study introduces Thresholding-based Evaluation Metrics in terms of Precision, Recall, and F-measure that not only assess the system’s reliability but also quantify the measure of trustworthiness of an XAI technique for a given classifier. Experiments are conducted on three benchmark datasets: CK+, RAFD, and RAF-DB, using classifiers including VGG19, ResNet18, GoogleNet, DenseNet121, and EfficientNet, and evaluated for different XAI techniques. The results demonstrate that the proposed metrics are as efficient as the traditional metrics and advance the assessment by quantifying the reliability measure, increasing their acceptability for human-centered applications. The source code will be made publicly available at <a href="https://github.com/Sayankumar007/FER-XAI-ThreshEvalMetrics">https://github.com/Sayankumar007/FER-XAI-ThreshEvalMetrics</a>.</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

Revealing the human-like similarities in automated facial expression recognition: an empirical investigation using eXplainable artificial intelligence

  • 期刊论文

  • 发表日期：

    2026年2月28日

查看全文
  • Sayan Kumar Bhowmick,
  • Asit Barman,
  • Swalpa Kumar Roy,
  • Paramartha Dutta

摘要

Human behavior analysis significantly depends on facial expression recognition, where deep learning has enabled the development of visionary models that can surpass human-level performance. Explainable Artificial Intelligence (XAI) techniques are employed to validate the trustworthiness of a trained convolutional neural network by providing interpretable heatmaps, generated using recent techniques, including GradCAM, GradCAM++, LayerCAM, and ScoreCAM. These saliency heatmaps highlight the critical facial regions used by the classifiers, thus aligning the system’s behavior with human cognitive processes. Metrics such as average drop, confidence increase, and win percentage are utilized to assess the system’s reliability by analyzing these heatmaps, but they can not quantify the measure of trustworthiness. This study introduces Thresholding-based Evaluation Metrics in terms of Precision, Recall, and F-measure that not only assess the system’s reliability but also quantify the measure of trustworthiness of an XAI technique for a given classifier. Experiments are conducted on three benchmark datasets: CK+, RAFD, and RAF-DB, using classifiers including VGG19, ResNet18, GoogleNet, DenseNet121, and EfficientNet, and evaluated for different XAI techniques. The results demonstrate that the proposed metrics are as efficient as the traditional metrics and advance the assessment by quantifying the reliability measure, increasing their acceptability for human-centered applications. The source code will be made publicly available at https://github.com/Sayankumar007/FER-XAI-ThreshEvalMetrics.