Scholarlink

In this comprehensive study, we explore the Neural Architectures over the food Food101 dataset. Ten state-of-the-art Convolutional Neural Network (CNN) architectures serve as the foundation of our exploration which includes VGG16 [1], VGG19 [1], GoogleNet [2], ResNet50 [3], DenseNet [4], MobileNet [5], EfficientB0 [6], EfficientB7 [6], inception-ResNet [7], and XceptionNet [8]. Initially, these architectures were trained through transfer learning without any augmentation and fine-tuning to observe the results over the food101 dataset. Subsequently, we applied augmentation techniques and fine-tuning in order to observe their impact on model performance. Through conscientious comparative analysis, we unveil insights into the strengths and weaknesses of each architecture under standardized conditions. Our work extends beyond the empirical, incorporating a quantitative assessment of model performance, visually compelling representations of key metrics, and a thorough exploration of hyper-parameter settings. This work thus stands as a holistic exploration of the nuances within the realm of food images, promising both depth and applicability in the field. This work also presents the comparison of results with the other works in which our fine-tuned model was able to defeat the accuracy of other models over the food101 Dataset.

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

Comparative Analysis of Deep State-of-the-Art CNN Architectures Over Food-101 Dataset Using Transfer Learning and the Impact of Fine Tuning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Faisal Rasheed,
  • Arshad Ahmad Dar,
  • Sarfaraz Ahmad,
  • Gousiya Hussain

摘要

In this comprehensive study, we explore the Neural Architectures over the food Food101 dataset. Ten state-of-the-art Convolutional Neural Network (CNN) architectures serve as the foundation of our exploration which includes VGG16 [1], VGG19 [1], GoogleNet [2], ResNet50 [3], DenseNet [4], MobileNet [5], EfficientB0 [6], EfficientB7 [6], inception-ResNet [7], and XceptionNet [8]. Initially, these architectures were trained through transfer learning without any augmentation and fine-tuning to observe the results over the food101 dataset. Subsequently, we applied augmentation techniques and fine-tuning in order to observe their impact on model performance. Through conscientious comparative analysis, we unveil insights into the strengths and weaknesses of each architecture under standardized conditions. Our work extends beyond the empirical, incorporating a quantitative assessment of model performance, visually compelling representations of key metrics, and a thorough exploration of hyper-parameter settings. This work thus stands as a holistic exploration of the nuances within the realm of food images, promising both depth and applicability in the field. This work also presents the comparison of results with the other works in which our fine-tuned model was able to defeat the accuracy of other models over the food101 Dataset.