Scholarlink

This paper introduces the Defiber layer, a Convolutional Neural Network (CNN) architecture component that systematically reduces input size during the convolution phase. In order to assess the feasibility of the proposed model, an empirical analysis was conducted using two image datasets: the CIFAR-10 dataset and an additional in-house dataset. Aiming to provide a comprehensive comparison, several existing CNN architectures are also included in this empirical analysis. The models are assessed based on accuracy and hardware performance criteria, including inference time and throughput. Our model surpasses expectations, outperforming some of the other evaluated models, both deep and mobile. Notably, the Defiber layer demonstrates remarkable effectiveness and consistent results during performance tests.

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

Proposing an Efficient CNN-Based Architecture for Image Processing

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Verner Rafael Ferreira,
  • Anne Magaly de Paula Canuto

摘要

This paper introduces the Defiber layer, a Convolutional Neural Network (CNN) architecture component that systematically reduces input size during the convolution phase. In order to assess the feasibility of the proposed model, an empirical analysis was conducted using two image datasets: the CIFAR-10 dataset and an additional in-house dataset. Aiming to provide a comprehensive comparison, several existing CNN architectures are also included in this empirical analysis. The models are assessed based on accuracy and hardware performance criteria, including inference time and throughput. Our model surpasses expectations, outperforming some of the other evaluated models, both deep and mobile. Notably, the Defiber layer demonstrates remarkable effectiveness and consistent results during performance tests.