Scholarlink

<p>The sudden upsurge in deployment of deep Convolutional Neural Networks (CNNs) for logo infringement detection has remarkably strengthened performance accuracy, yet their opaque decision-making process limits trustworthiness. This further hinders their interpretability and reduces their reliability in real-world situations. Therefore, this article builds upon the work of Gupta et al. on logo infringement detection employing a ResNet classifier and targets model explainability. This article utilises Gradient-weighted Class Activation Mapping (Grad-CAM) on a fine-tuned ResNet-101 model, generating heatmaps that facilitate model interpretation and explanation. This visualisation demarcates the most discriminative image regions that supervise the model’s prediction. The Grad-CAM heatmaps illustrate that the model distinctly targets the salient logo features, ignoring the irrelevant ones. Furthermore, quantitative evaluation metrics such as Area Ratio, Energy Concentration and Intersection over Union (IoU) were calculated for the resultant heatmaps, which further demonstrate the efficacy of Grad-CAM.</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

Enhancing Explainability Through Grad-CAM Lens in Logo Infringement Detection

  • 期刊论文

  • 发表日期：

    2026年4月9日

查看全文
  • Pratishtha Gupta,
  • Bhawna Narwal,
  • A.K. Mohapatra

摘要

The sudden upsurge in deployment of deep Convolutional Neural Networks (CNNs) for logo infringement detection has remarkably strengthened performance accuracy, yet their opaque decision-making process limits trustworthiness. This further hinders their interpretability and reduces their reliability in real-world situations. Therefore, this article builds upon the work of Gupta et al. on logo infringement detection employing a ResNet classifier and targets model explainability. This article utilises Gradient-weighted Class Activation Mapping (Grad-CAM) on a fine-tuned ResNet-101 model, generating heatmaps that facilitate model interpretation and explanation. This visualisation demarcates the most discriminative image regions that supervise the model’s prediction. The Grad-CAM heatmaps illustrate that the model distinctly targets the salient logo features, ignoring the irrelevant ones. Furthermore, quantitative evaluation metrics such as Area Ratio, Energy Concentration and Intersection over Union (IoU) were calculated for the resultant heatmaps, which further demonstrate the efficacy of Grad-CAM.