A convolutional neural network based on optimized structure and its lightweighting

Jinyong Deng; Zhiheng Zhao; Yang Liu; Yongzhe Chen; Zhefan Zhang; Yu Zhang

doi:10.1117/12.2659679

22 November 2022 A convolutional neural network based on optimized structure and its lightweighting

Jinyong Deng, Zhiheng Zhao, Yang Liu, Yongzhe Chen, Zhefan Zhang, Yu Zhang

Proceedings Volume 12475, Second International Conference on Advanced Algorithms and Signal Image Processing (AASIP 2022); 1247513 (2022) https://doi.org/10.1117/12.2659679
Event: Second International Conference on Advanced Algorithms and Signal Image Processing (AASIP 2022), 2022, Hulun Buir, China

Abstract

In this paper, we design a convolutional neural network based on the ideas of depthwise separable convolution and inverted residual module. The scaling factor of BN layer is used as a measure for channel pruning of the network model to compress it. By analyzing the layer-by-layer pruning process of conventional convolution, the layer-by-layer pruning method with depthwise separable convolution and inverted residual structure is proposed to prune the channels of the network model, and finally, the channel pruning strategy of classification simplification network is developed. Tests on the selected dataset showed that the classification accuracy of the pruned and fine-tuned network model is 97.7% when the pruning rate is 0.7.

Citation Download Citation

Jinyong Deng, Zhiheng Zhao, Yang Liu, Yongzhe Chen, Zhefan Zhang, and Yu Zhang "A convolutional neural network based on optimized structure and its lightweighting", Proc. SPIE 12475, Second International Conference on Advanced Algorithms and Signal Image Processing (AASIP 2022), 1247513 (22 November 2022); https://doi.org/10.1117/12.2659679

ACCESS THE FULL ARTICLE

INSTITUTIONAL
Select your institution to access the SPIE Digital Library.

SELECT YOUR INSTITUTION

PERSONAL
Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.

PERSONAL SIGN IN

No SPIE Account? Create one

PURCHASE THIS CONTENT

SUBSCRIBE TO DIGITAL LIBRARY

50 downloads per 1-year subscription

Members: $195

Non-members: $335 ADD TO CART

25 downloads per 1 - year subscription

Members: $145

Non-members: $250 ADD TO CART

PURCHASE SINGLE ARTICLE

Includes PDF, HTML & Video, when available