Scholarlink

Data acquisition and labelling at scale can be time-consuming, expensive, and often infeasible, especially in resource-constrained environments. This has driven the search for methods that can reduce the dependency on labelled data while still maintaining high performance. Transfer learning is one promising solution to this problem. It allows for faster convergence and better performance with significantly less data. Hence, this paper explores the effect of custom architectural changes like Global Average Pooling (GAP), Average Depth Constraints, and Depth Constraints in improving transfer learning for image classification. Five state-of-the-art convolutional neural networks—DenseNet169, EfficientNetB0, MobileNetV2, XceptionNet, and ResNet50—were tested on the Intel Image Dataset (simple, six classes) and the MIT Indoor Dataset (complex, 67 classes) to analyse their adaptability towards datasets of varying complexity. The results indicate that models including these custom layers consistently outperform the standard architectures, with the MobileNetV2 obtaining a maximum accuracy of ~ 90.0% on the Intel Dataset and XceptionNet showing remarkable generalization as it holds 71.6% accuracy on the more complex MIT Indoor Dataset. The results also demonstrate the effectiveness of MobileNetV2 and EfficientNetB0 in low-resource settings, offering high performance at low parameter counts. The proposed modifications therefore address the challenges of limited labelled data and computational resources to open new avenues in applications such as healthcare, security, and autonomous systems for efficient transfer learning.

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

Advancing Transfer Learning with GAP, Depth Constraints, and Average Depth Constraints for Diverse Datasets

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Kartik Garg,
  • Reena Gupta

摘要

Data acquisition and labelling at scale can be time-consuming, expensive, and often infeasible, especially in resource-constrained environments. This has driven the search for methods that can reduce the dependency on labelled data while still maintaining high performance. Transfer learning is one promising solution to this problem. It allows for faster convergence and better performance with significantly less data. Hence, this paper explores the effect of custom architectural changes like Global Average Pooling (GAP), Average Depth Constraints, and Depth Constraints in improving transfer learning for image classification. Five state-of-the-art convolutional neural networks—DenseNet169, EfficientNetB0, MobileNetV2, XceptionNet, and ResNet50—were tested on the Intel Image Dataset (simple, six classes) and the MIT Indoor Dataset (complex, 67 classes) to analyse their adaptability towards datasets of varying complexity. The results indicate that models including these custom layers consistently outperform the standard architectures, with the MobileNetV2 obtaining a maximum accuracy of ~ 90.0% on the Intel Dataset and XceptionNet showing remarkable generalization as it holds 71.6% accuracy on the more complex MIT Indoor Dataset. The results also demonstrate the effectiveness of MobileNetV2 and EfficientNetB0 in low-resource settings, offering high performance at low parameter counts. The proposed modifications therefore address the challenges of limited labelled data and computational resources to open new avenues in applications such as healthcare, security, and autonomous systems for efficient transfer learning.