Scholarlink

<p>In recent years, efficient convolutional neural network (CNN) models have received limited attention due to their relatively inferior performance compared to other popular efficient models. In this paper, we propose a novel efficient CNN-based model, dubbed MoConv, which retains many advantages of CNNs. Our approach begins with an analysis of the Transformer encoder structure, leading to the Comprehensive Convolution Attention (CCA) mechanism. Using CCA with the Inverted Residual Block (IRB), we designed the fundamental MoConv block analogous to the Transformer encoder and constructed the MoConv model for various computer vision tasks. The effectiveness of MoConv is demonstrated through benchmarks like ImageNet, MS COCO, and ADE20K. For instance, MoConv-S/T/N models, trained with one or two RTX4090 GPUs on ImageNet, achieve top-1 accuracy rates of 78.6%, 75.7%, and 72.2%, respectively, on ImageNet-1k, with computational overheads of only 5.6M/1.4GFLOPs, 2.3M/0.5GFLOPs, and 1.2M/0.3GFLOPs. This clearly surpasses the performance of SwiftFormer, ConvMLP, EMO, MobileViG, and EfficientFormer, which are among the most outstanding performers in the same class. Consequently, our models stand out as excellent choices among mainstream lightweight models.</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

Rejuvenating efficient convolutional neural networks

  • 期刊论文

  • 发表日期：

    2025年5月23日

查看全文
  • Wei Xu,
  • Yi Wan,
  • Kaiyuan Wang

摘要

In recent years, efficient convolutional neural network (CNN) models have received limited attention due to their relatively inferior performance compared to other popular efficient models. In this paper, we propose a novel efficient CNN-based model, dubbed MoConv, which retains many advantages of CNNs. Our approach begins with an analysis of the Transformer encoder structure, leading to the Comprehensive Convolution Attention (CCA) mechanism. Using CCA with the Inverted Residual Block (IRB), we designed the fundamental MoConv block analogous to the Transformer encoder and constructed the MoConv model for various computer vision tasks. The effectiveness of MoConv is demonstrated through benchmarks like ImageNet, MS COCO, and ADE20K. For instance, MoConv-S/T/N models, trained with one or two RTX4090 GPUs on ImageNet, achieve top-1 accuracy rates of 78.6%, 75.7%, and 72.2%, respectively, on ImageNet-1k, with computational overheads of only 5.6M/1.4GFLOPs, 2.3M/0.5GFLOPs, and 1.2M/0.3GFLOPs. This clearly surpasses the performance of SwiftFormer, ConvMLP, EMO, MobileViG, and EfficientFormer, which are among the most outstanding performers in the same class. Consequently, our models stand out as excellent choices among mainstream lightweight models.