A light-weight model with granularity feature representation for fine-grained visual classification
by Qiumei Zheng; Tianqi Peng; Ding Huang; Fenghua Wang; Nengxiang Xu
International Journal of Computational Science and Engineering (IJCSE), Vol. 27, No. 3, 2024

Abstract: Fine-grained image recognition can provide a more precise recognition technique for industrial production and applications. However, since it is difficult to capture comprehensive features and discriminative regions in convolutional neural networks (CNN), this ability is largely limited. With a lightweight orientation, we here use the advantage of Transformer in capturing global features by combining the technically mature CNN, and propose a lightweight model MV-GFR based on MobileViT. Further, we also propose three lightweight modules to help the network capture more subtle differences. First, we used the training module to provide the network with richer granularity information while ensuring its global integrity. Second, we used the feature part mask module in combining the diversity of CNN and the saliency of the transformer. Finally, we used the feature fusion module to integrate features of different levels and generate a complement between the global and local features. We then demonstrated the effectiveness of this scheme through experiments on three commonly used datasets.

Online publication date: Fri, 03-May-2024

The full text of this article is only available to individual subscribers or to users at subscribing institutions.

 
Existing subscribers:
Go to Inderscience Online Journals to access the Full Text of this article.

Pay per view:
If you are not a subscriber and you just want to read the full contents of this article, buy online access here.

Complimentary Subscribers, Editors or Members of the Editorial Board of the International Journal of Computational Science and Engineering (IJCSE):
Login with your Inderscience username and password:

Username:        Password:          Forgotten your password?

Want to subscribe?
A subscription gives you complete access to all articles in the current issue, as well as to all articles in the previous three years (where applicable). See our Orders page to subscribe.

If you still need assistance, please email subs@inderscience.com