Scholarlink

<p>Object recognition plays a critical role in various real-world applications, from intelligent surveillance to autonomous systems. This study presents a comparative evaluation of deep convolutional neural network (CNN)-based feature extraction techniques—ResNet50, Xception, and VGG19—combined with four classical classification algorithms: Gaussian Naïve Bayes, k-Nearest Neighbors (k-NN), Decision Tree, and Random Forest. Using the Caltech-101 dataset, which comprises 8677 images across 101 object categories, we assess each hybrid combination using recognition accuracy, Area Under the Curve (AUC), Root Mean Square Error (RMSE), and processing time. The results demonstrate that the combination of VGG19 with Random Forest achieves the highest recognition accuracy (91.91%) and the lowest RMSE (16.73%), while Xception with Random Forest attains the highest AUC (95.27%). The study also includes statistical validation using paired t-tests, confirming the significance of the performance differences (<i>p</i> &lt; 0.05). These findings underscore the potential of hybrid CNN-classifier models for effective and computationally efficient object recognition. The proposed framework provides valuable insights into selecting lightweight and interpretable model architectures, particularly for deployment in embedded or resource-constrained environments.</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

Performance Comparison of Deep Feature Extraction and Classification Techniques in Object Recognition

  • 期刊论文

  • 发表日期：

    2025年12月24日

查看全文
  • Muhammad Irsyad Abdullah,
  • Ahmed Alkhayyat,
  • Gadug Sudhamsu,
  • Pooja Rani,
  • Aman Shankhyan,
  • Jasgurpreet Singh Chohan,
  • M. Janaki Ramudu,
  • Devendra Singh

摘要

Object recognition plays a critical role in various real-world applications, from intelligent surveillance to autonomous systems. This study presents a comparative evaluation of deep convolutional neural network (CNN)-based feature extraction techniques—ResNet50, Xception, and VGG19—combined with four classical classification algorithms: Gaussian Naïve Bayes, k-Nearest Neighbors (k-NN), Decision Tree, and Random Forest. Using the Caltech-101 dataset, which comprises 8677 images across 101 object categories, we assess each hybrid combination using recognition accuracy, Area Under the Curve (AUC), Root Mean Square Error (RMSE), and processing time. The results demonstrate that the combination of VGG19 with Random Forest achieves the highest recognition accuracy (91.91%) and the lowest RMSE (16.73%), while Xception with Random Forest attains the highest AUC (95.27%). The study also includes statistical validation using paired t-tests, confirming the significance of the performance differences (p < 0.05). These findings underscore the potential of hybrid CNN-classifier models for effective and computationally efficient object recognition. The proposed framework provides valuable insights into selecting lightweight and interpretable model architectures, particularly for deployment in embedded or resource-constrained environments.