Hardware-Friendly and Efficient Vision Transformer for Deployment on Low-Power Embedded Device

The Transformer architecture has achieved remarkable success across numerous computer vision tasks due to its superior capability for global dependency modeling. However, the high computational complexity and hardware-unfriendly operations such as Layer Normalization (LN), Softmax, and GELU severely hinder its deployment on resource-constrained platforms. To address these challenges, this paper proposes a hardware-friendly CNN-Transformer hybrid pyramid architecture that effectively balances accuracy, efficiency, and deployability. The proposed model integrates convolutional bottlenecks with Transformer encoders to capture both local and global contextual information while maintaining low computational cost. A pyramid feature extraction structure is further introduced to enhance multi-scale semantic representation. To improve hardware efficiency, we redesign key nonlinear components by introducing hardware-friendly activation, normalization, and Softmax approximations. Specifically, GELU and LN are replaced by ReLU and Batch Normalization (BN), and a simplified logarithmic-exponential formulation termed Softmax2 is proposed, which eliminates complex exponential and division operations, significantly reducing hardware implementation cost. Extensive experiments demonstrate the effectiveness of the proposed framework. The experimental results validate that the proposed architecture offers a promising and practical solution for real-time and embedded vision applications.