A Novel Fusion Pruning-Processed Lightweight CNN for Local Object Recognition on Resource-Constrained Devices

In recent years, smartphones and smart educational products have become important catalysts for the development of consumer electronics. However, deploying large convolutional neural networks (CNNs) models on resource-constrained devices like smartphones is impractical due to the lack of high-perfor...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on consumer electronics 2024-10, p.1-1
Main Authors: Xie, Can, Zhai, Xiaobing, Chi, Haiyang, Li, Wenxiang, Li, Xiaolin, Sha, Yuyang, Li, Kefeng
Format: Article
Language:English
Subjects:
Accuracy
Computational modeling
Convolutional neural networks
Fusion pruning
Image recognition
Lightweight
Local object recognition
Mobile devices
Neural networks
Object recognition
Performance evaluation
Quantization (signal)
Smart phones
Training
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In recent years, smartphones and smart educational products have become important catalysts for the development of consumer electronics. However, deploying large convolutional neural networks (CNNs) models on resource-constrained devices like smartphones is impractical due to the lack of high-performance central processing units (CPUs), graphics processing units (GPUs), and large-capacity memory. To address these challenges, we propose a new Fusion Pruning (FP) method aimed at compressing and accelerating large CNN models by leveraging L2 norm and equivalent transformation of the receptive field. To validate the feasibility of the proposed method in language education, we have developed an application deployed on smartphones. This application enables offline object recognition without the need for additional hardware platforms like Jetson Xavier NX. Additionally, we have also explored the performance of the FP method in the field of household robotics, obtaining satisfactory results. Experimental results demonstrate that the proposed method can accomplish tasks on resource-constrained devices.
ISSN:0098-3063
1558-4127
DOI:10.1109/TCE.2024.3475517