Year: 2024
Type: article
Abstract: We focus on developing a lightweight model for resource-constrained devices, building on MobileViT, a hybrid model that combines the strengths of Transformers and CNNs to balance high accuracy and computational efficiency for image classification. Transformers, while effective at capturing global information, often have higher computational costs than CNNs due to the complexity of their self-attention mechanism. To address this, we introduce the Token Merging (ToMe) technique into MobileViT to reduce computational costs. However, because the number of tokens changes during merging, ToMe cannot be directly applied to MobileViT without adjustments. We propose simple methods, specifically reshaping features and removing skip connections, to resolve this issue. Additionally, we make adjustments to MobileViT’s structure to better support the application of ToMe. Our approach improves inference efficiency while retaining a competitive level of accuracy. The resulting models achieve a balance between performance and computational speed, offering a practical solution for hybrid architectures. This work shows the potential of ToMe-based techniques to broaden the range of lightweight model options, catering to diverse application requirements.
Source: Electronics
Institution: Osaka Metropolitan University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,728
OpenAlex ID: W4405642006