Unpaired Multisource Imagery Joint Learning via Lightweight Network for Tread Tire Recognition

Tread tire recognition has attracted much attention during the last decade since its vital practival application value in criminal investigation cases, but the feature representation from only one source imagery of tire pattern can not reflect its semantic attributes because of some unknown interfer...

Full description

Saved in:
Bibliographic Details
Published in2024 6th International Conference on Natural Language Processing (ICNLP) pp. 585 - 591
Main Authors Zhong, Yulu, Wang, Jiasen, Fang, Jie
Format Conference Proceeding
LanguageEnglish
Published IEEE 22.03.2024
Subjects
CycleGAN
Image recognition
lightweight network
Natural language processing
Noise
Pattern recognition
perception retrieval
Semantics
Testing
Tires
Training
Tread tire recognition
unpaired multisource imagery
Online AccessGet full text

Cover

More Information
Summary:Tread tire recognition has attracted much attention during the last decade since its vital practival application value in criminal investigation cases, but the feature representation from only one source imagery of tire pattern can not reflect its semantic attributes because of some unknown interfere noises, and hence influence the recognition performance. To address the aforementioned problem, this work presents two novel unpaired tire pattern imagery recognition frameworks, including a multisource joint learning framework and a cross-domain expansion learning framework, the former needs multisource images in training phase and testing phases while the latter needs multisource images in training phase and only one source images in testing phase. Specifically, the multisource joint learning framework utilizes a perception retrieval based fusion module to enhance the complementary information of surface pattern and tire mark to each other, then uses a two-stream lightweight network based on GhostNet to exploit the latent semantic information of multisource images and finalize the recognition. The cross-domain expansion learning framework uses a CycleGAN to generate corresponding image series of unavailable domain according to ones in existing domain, and utilizes the same two-stream classification network in multisource joint learning framework to finalize the recognition. Finally, the proposed two frameworks achieve competitive performances on the challenging CIIP-TPID dataset, which demonstrate their superiorities.
DOI:10.1109/ICNLP60986.2024.10692996