Scholarlink

<p>GldReLU, as a novel activation function, is proposed in this research work. The number of activation functions have been investigated for Deep Learning (DL) models to improve the performance. The Rectified Linear Unit (ReLU) is a widely used activation function. Although a number of alternatives to ReLU have been used to enhance training performance and stability and to understand the ways in which ReLU interacts with different optimization strategies, weight initialization techniques, and network architectures, we propose a novel activation function, GldReLU. By scaling the ReLU function with the Golden Ratio phi (φ). The ReLU can be replaced with GldReLU in DL models. The proposed GldReLU is investigated with Residual Network (RESNET50), Visual Geometry Group (VGG16) and a customized Artificial Neural Network (ANN) model, and the outcomes are compared with every ReLU variant. The experiment is conducted with 118 real-time Chest X-ray images and 10,166 images from Kaggle for multi-classification. Furthermore more the experiment is conducted with 225 Teeth root X-ray images for binary classification. The Caltech-101 dataset is used and the results are compared with ReLU and GldReLU. Then for the text data Pima Indian Diabetes Dataset has been utilized. The comparison is done with the benchmark dataset, Cifar10. The findings demonstrated that accuracy, GldReLU outperforms ReLU and its variants.Cifar10 dataset with RESNET50 was 83% of accuracy and with modified RESNET50, is of 87%.The accuracy for the classification of real-time Chest X-ray images in RESNET50 is 85% and in modified RESNET50 is 89%. The Caltech-101 dataset is also taken for evaluation. The test accuracy improvement was 10% in GldReLU when compared with ReLU in the network. Besides this, the accuracy of the customized Artificial Neural Network (ANN) for the Pima Indian Diabetes Dataset with four hidden layers and ReLU AF is 71%, while the accuracy with GldReLU is 82%. Thus, the proposed novel GldReLU outperforms in the model with all datasets.</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

A novel GldReLU activation function with enhanced RESNET50 for classification of X-ray images

  • 期刊论文

  • 发表日期：

    2025年8月4日

查看全文
  • P. Pankaja Lakshmi,
  • M. Sivagami

摘要

GldReLU, as a novel activation function, is proposed in this research work. The number of activation functions have been investigated for Deep Learning (DL) models to improve the performance. The Rectified Linear Unit (ReLU) is a widely used activation function. Although a number of alternatives to ReLU have been used to enhance training performance and stability and to understand the ways in which ReLU interacts with different optimization strategies, weight initialization techniques, and network architectures, we propose a novel activation function, GldReLU. By scaling the ReLU function with the Golden Ratio phi (φ). The ReLU can be replaced with GldReLU in DL models. The proposed GldReLU is investigated with Residual Network (RESNET50), Visual Geometry Group (VGG16) and a customized Artificial Neural Network (ANN) model, and the outcomes are compared with every ReLU variant. The experiment is conducted with 118 real-time Chest X-ray images and 10,166 images from Kaggle for multi-classification. Furthermore more the experiment is conducted with 225 Teeth root X-ray images for binary classification. The Caltech-101 dataset is used and the results are compared with ReLU and GldReLU. Then for the text data Pima Indian Diabetes Dataset has been utilized. The comparison is done with the benchmark dataset, Cifar10. The findings demonstrated that accuracy, GldReLU outperforms ReLU and its variants.Cifar10 dataset with RESNET50 was 83% of accuracy and with modified RESNET50, is of 87%.The accuracy for the classification of real-time Chest X-ray images in RESNET50 is 85% and in modified RESNET50 is 89%. The Caltech-101 dataset is also taken for evaluation. The test accuracy improvement was 10% in GldReLU when compared with ReLU in the network. Besides this, the accuracy of the customized Artificial Neural Network (ANN) for the Pima Indian Diabetes Dataset with four hidden layers and ReLU AF is 71%, while the accuracy with GldReLU is 82%. Thus, the proposed novel GldReLU outperforms in the model with all datasets.