Scholarlink

Convolutional Neural Networks (CNNs) have revolutionized computer vision tasks by automatically learning hierarchical representations from raw input data. Despite their success, CNNs struggle with effectively capturing contextual information across different dimensions within the network. Attention mechanisms, particularly the Convolutional Block Attention Module (CBAM), have emerged as a solution to enhance CNNs’ performance by enabling them to focus on relevant regions and features while suppressing irrelevant ones. In this paper, we investigate the optimal placements of the CBAM module within the ResNet-18 architecture, a widely used CNN architecture. Our experiments conducted on three diverse datasets—MIT Indoor Scenes Dataset, CIFAR-10 dataset, and MNIST dataset—aim to explore the impact of CBAM placement on model performance. We address key research questions regarding the influence of CBAM placement on accuracy, convergence speed, and the model’s ability to capture contextual information. Through rigorous experimentation and analysis, we provide insights into the integration of attention mechanisms within CNN architectures, contributing to advancements in computer vision tasks. The codes of this paper can be found here: https://github.com/ARMIS2111/Investigation-on-Placements-of-CBAM-Module-in-ResNet-18-Architecture .

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

Investigation on Placement of Convolutional Block Attention Module in ResNet-18 Architecture

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Archit Mishra,
  • Mainak Bandyopadhyay,
  • Priyanshu Priyadarshi

摘要

Convolutional Neural Networks (CNNs) have revolutionized computer vision tasks by automatically learning hierarchical representations from raw input data. Despite their success, CNNs struggle with effectively capturing contextual information across different dimensions within the network. Attention mechanisms, particularly the Convolutional Block Attention Module (CBAM), have emerged as a solution to enhance CNNs’ performance by enabling them to focus on relevant regions and features while suppressing irrelevant ones. In this paper, we investigate the optimal placements of the CBAM module within the ResNet-18 architecture, a widely used CNN architecture. Our experiments conducted on three diverse datasets—MIT Indoor Scenes Dataset, CIFAR-10 dataset, and MNIST dataset—aim to explore the impact of CBAM placement on model performance. We address key research questions regarding the influence of CBAM placement on accuracy, convergence speed, and the model’s ability to capture contextual information. Through rigorous experimentation and analysis, we provide insights into the integration of attention mechanisms within CNN architectures, contributing to advancements in computer vision tasks. The codes of this paper can be found here: https://github.com/ARMIS2111/Investigation-on-Placements-of-CBAM-Module-in-ResNet-18-Architecture .