Article: A transformer-based model of fine-grained image classification for cigarette trademark identification Journal: International Journal of Advanced Mechatronic Systems (IJAMECHS) 2026 Vol.13 No.2 pp.122 - 132 Abstract: Traditional methods in fine-grained image recognition for cigarette trademarks have a limitation of low accuracy. In this paper, we propose a fine-grained image classification model with a transformer-based network architecture for cigarette images. In the proposed model, the local feature refinement module focuses on the critical areas of cigarette trademarks and packs via channel and spatial attention mechanisms, while the adaptive feature fusion module integrates multiscale features to enhance classification accuracy. Additionally, we design a novel loss function based on weighted cross-entropy and region sensitivity, allowing the model to focus on both global and local fine-grained features. Experimental results demonstrate that our proposed model achieves an accuracy of 99.5% on the Furongwang dataset, 98.1% on the Yuxi dataset, and 98.8% on the Liqun dataset, surpassing the best-performing existing method by up to 1.1%. These results confirm the effectiveness of our approach in the fine-grained classification of cigarette trademarks. Inderscience Publishers - linking academia, business and industry through research

Title: A transformer-based model of fine-grained image classification for cigarette trademark identification

Authors: Xiaohui Li; Haoran Zhu; Yupeng Xu; Changshen Yan; Lan Yao; Feng Zeng

Addresses: National Tobacco Quality Supervision and Test Center, Zhengzhou, China ' China Tobacco Guangxi Industrial Co., Ltd., Nanning, China ' National Tobacco Quality Supervision and Test Center, Zhengzhou, China ' School of Computer Science and Engineering, Central South University, Changsha, China; School of Mathematics, Hunan University, Changsha, China ' School of Computer Science and Engineering, Central South University, Changsha, China; School of Mathematics, Hunan University, Changsha, China ' School of Computer Science and Engineering, Central South University, Changsha, China; School of Mathematics, Hunan University, Changsha, China

Abstract: Traditional methods in fine-grained image recognition for cigarette trademarks have a limitation of low accuracy. In this paper, we propose a fine-grained image classification model with a transformer-based network architecture for cigarette images. In the proposed model, the local feature refinement module focuses on the critical areas of cigarette trademarks and packs via channel and spatial attention mechanisms, while the adaptive feature fusion module integrates multiscale features to enhance classification accuracy. Additionally, we design a novel loss function based on weighted cross-entropy and region sensitivity, allowing the model to focus on both global and local fine-grained features. Experimental results demonstrate that our proposed model achieves an accuracy of 99.5% on the Furongwang dataset, 98.1% on the Yuxi dataset, and 98.8% on the Liqun dataset, surpassing the best-performing existing method by up to 1.1%. These results confirm the effectiveness of our approach in the fine-grained classification of cigarette trademarks.

Keywords: fine-grained image recognition; transformer; attention mechanism; feature fusion.

DOI: 10.1504/IJAMECHS.2026.153232

International Journal of Advanced Mechatronic Systems, 2026 Vol.13 No.2, pp.122 - 132

Received: 02 Jan 2025
Accepted: 17 Mar 2025
Published online: 29 Apr 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article

Title: A transformer-based model of fine-grained image classification for cigarette trademark identification

Keep up-to-date