摘要
为了提升传统CNN的远距离特征提取能力,将深度可分离卷积操作嵌入到改进MetaFormer架构中,添加了通道混洗操作,提出了一种创新的轻量化CNN网络模型:ViTNet。通过这种融合策略,不仅保留了ViTs的灵活性和扩展性,还增强了CNN模型的图像特征提取能力。ViTNet的轻量化优势,使其能够高效运行于计算资源有限的设备上。实验结果表明:在cifar10上,ViTNet-1.0×比MobileNetV2提高了1.8%准确率,延迟降低了32%,表现出良好的竞争力。
In order to improve the remote feature extraction capability of traditional CNN,this paper presents an innovative lightweight CNN network model known as ViTNet,with the deep detachable convolution operation embedded into the improved MetaFormer architecture and with channel shuffle operation added.This fusion strategy not only maintains the flexibility and scalability of ViTs,but also enhances the image feature extraction capability of the CNN model.Furthermore,lightweight enables ViTNet to operate efficiently on devices with restricted computing resources.The experimental results show that on cifar10,ViTNet-1.0×is more competitive than MobileNetV2 with accuracy improved by 1.8%and latency reduced by 32%.
作者
徐飞
禹婷婷
张乐怡
张瑞轩
XU Fei;YU Tingting;ZHANG Leyi;ZHANG Ruixuan(School of Computer Science and Engineering,Xi’an Technological University,Xi'an 710021,China)
出处
《西安工业大学学报》
CAS
2024年第4期506-513,共8页
Journal of Xi’an Technological University
基金
陕西省科技厅区域创新能力指导计划项目(2022QFY01-14)
咸阳市科技局重点研发项目(2021ZDYF-NY-0019)。
关键词
图像识别
深度可分离卷积
轻量级网络
模型压缩
image identification
depth-wise separable convolution
lightweight network
model compression
作者简介
徐飞(1980-),男,副教授,主要研究方向为边缘计算、自适应中间件、企业信息化,E-mail:29112462@qq.com。
