Scholarlink

<p>In today’s landscape of human activity recognition (HAR), many existing models are computationally intensive, leading to challenges in real-world deployment. Moreover, there is a notable lack of attention mechanism-based models that effectively balance performance with efficiency. To address these issues, we propose a lightweight model, DIAT-Separable-CNN-ECA-HARNet, which employs a dense block structure along with separable convolution layers. This approach not only significantly reduces the number of parameters while ensuring computational efficiency but also enhances feature reuse and improves gradient flow. Additionally, we integrate the efficient channel attention (ECA) mechanism, enabling the model to focus on critical features in the input data. Furthermore, we utilize explainable artificial intelligence (XAI) techniques, which enhance model interpretability by providing insights into decision-making processes. This transparency helps stakeholders trust and understand model predictions. The model’s performance is rigorously evaluated on a seven-class human activity dataset, achieving an impressive accuracy of 99.34 %. Further testing on the DIAT-<InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="11760_2025_3838_Article_IEq1.gif" Format="GIF" Height="12" Rendition="HTML" Resolution="72" Type="Linedraw" Width="15" /> </InlineMediaObject> <EquationSource Format="TEX">\(\mu \)</EquationSource> <EquationSource Format="MATHML"><math> <mi>μ</mi> </math></EquationSource> </InlineEquation>RadHAR dataset demonstrates an even higher accuracy of 99.65 %. Notably, the proposed model requires only 0.33 giga (G) floating point operations (FLOPs) and consists of merely 0.12 million (M) parameters, underscoring its efficiency. Overall, the DIAT-Separable-CNN-ECA-HARNet model not only outperforms existing models and methods but also presents a promising solution for real-world applications in the evolving field of human activity recognition technology, paving the way for advancements in efficient and effective systems. For code and further details, please visit: <a href="https://github.com/ajaywagh007/DIAT-Separable-CNN-ECA-HARNet-A-Lightweight-and-Explainable-Model-for-Efficient-HAR/tree/main">https://github.com/ajaywagh007/DIAT-Separable-CNN-ECA-HARNet-A-Lightweight-and-Explainable-Model-for-Efficient-HAR/tree/main</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

Diat-separable-cnn-eca-harnet: a lightweight and explainable model for efficient human activity recognition

  • 期刊论文

  • 发表日期：

    2025年1月28日

查看全文
  • Ajay Waghumbare,
  • Upasna Singh

摘要

In today’s landscape of human activity recognition (HAR), many existing models are computationally intensive, leading to challenges in real-world deployment. Moreover, there is a notable lack of attention mechanism-based models that effectively balance performance with efficiency. To address these issues, we propose a lightweight model, DIAT-Separable-CNN-ECA-HARNet, which employs a dense block structure along with separable convolution layers. This approach not only significantly reduces the number of parameters while ensuring computational efficiency but also enhances feature reuse and improves gradient flow. Additionally, we integrate the efficient channel attention (ECA) mechanism, enabling the model to focus on critical features in the input data. Furthermore, we utilize explainable artificial intelligence (XAI) techniques, which enhance model interpretability by providing insights into decision-making processes. This transparency helps stakeholders trust and understand model predictions. The model’s performance is rigorously evaluated on a seven-class human activity dataset, achieving an impressive accuracy of 99.34 %. Further testing on the DIAT- \(\mu \) μ RadHAR dataset demonstrates an even higher accuracy of 99.65 %. Notably, the proposed model requires only 0.33 giga (G) floating point operations (FLOPs) and consists of merely 0.12 million (M) parameters, underscoring its efficiency. Overall, the DIAT-Separable-CNN-ECA-HARNet model not only outperforms existing models and methods but also presents a promising solution for real-world applications in the evolving field of human activity recognition technology, paving the way for advancements in efficient and effective systems. For code and further details, please visit: https://github.com/ajaywagh007/DIAT-Separable-CNN-ECA-HARNet-A-Lightweight-and-Explainable-Model-for-Efficient-HAR/tree/main