Scholarlink

<p>This work presents a comprehensive empirical evaluation of neural network configurations for image classification, with a specific focus on compressed domain processing applications. We systematically evaluated 406 distinct configurations across two architectures: a baseline dense-layer model (SimpleDense) and a compact hybrid convolutional network (HybridConvNet) that integrates established architectural components. Our study examined activation functions, network depths, learning rates, batch sizes, and optimizers using CIFAR-10 as a primary dataset, with additional validation on the full CIFAR-100 and JPEG-compressed data. SimpleDense achieved a peak validation accuracy of <b>0.5120</b>, while HybridConvNet reached <b>0.6900</b> on uncompressed data and maintained robust performance on moderately compressed images (JPEG quality 50) with an accuracy of <b>0.6486</b>. Comprehensive ablation studies revealed that multi-head self-attention contributed a <b>9.1</b> absolute accuracy improvement, while depthwise separable convolutions provided a 12.3% parameter reduction. Efficiency analysis showed HybridConvNet requires 2.1M parameters and a 15.2ms inference time, making it suitable for edge deployment. Comparisons with lightweight models demonstrate competitive accuracy-efficiency trade-offs, positioning this work as an empirical systematization that provides actionable insights for developing efficient neural networks in resource-constrained environments.</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

Compact HybridConvNet: an empirical study of convolutional architectures for image classification with applications to compressed domains

  • 期刊论文

  • 发表日期：

    2026年2月3日

查看全文
  • Naga Sirisha Rayala,
  • Veeranjaneyulu Naralasetti

摘要

This work presents a comprehensive empirical evaluation of neural network configurations for image classification, with a specific focus on compressed domain processing applications. We systematically evaluated 406 distinct configurations across two architectures: a baseline dense-layer model (SimpleDense) and a compact hybrid convolutional network (HybridConvNet) that integrates established architectural components. Our study examined activation functions, network depths, learning rates, batch sizes, and optimizers using CIFAR-10 as a primary dataset, with additional validation on the full CIFAR-100 and JPEG-compressed data. SimpleDense achieved a peak validation accuracy of 0.5120, while HybridConvNet reached 0.6900 on uncompressed data and maintained robust performance on moderately compressed images (JPEG quality 50) with an accuracy of 0.6486. Comprehensive ablation studies revealed that multi-head self-attention contributed a 9.1 absolute accuracy improvement, while depthwise separable convolutions provided a 12.3% parameter reduction. Efficiency analysis showed HybridConvNet requires 2.1M parameters and a 15.2ms inference time, making it suitable for edge deployment. Comparisons with lightweight models demonstrate competitive accuracy-efficiency trade-offs, positioning this work as an empirical systematization that provides actionable insights for developing efficient neural networks in resource-constrained environments.