EdgeNeXt: Efficiently Amalgamated CNN-Transformer Architecture for Mobile Vision Applications

In the pursuit of achieving ever-increasing accuracy, large and complex neural networks are usually developed. Such models demand high computational resources therefore cannot be deployed on edge devices. It is great interest to build resource-efficient general purpose due their usefulness in several application areas. this work, we strive effectively combine strengths both CNN Transformer propose a new efficient hybrid architecture EdgeNeXt. Specifically EdgeNeXt, introduce split depth-wise transpose attention (STDA) encoder that splits input tensors into multiple channel groups utilizes convolution along with self-attention across dimensions implicitly increase receptive field encode multi-scale features. Our extensive experiments classification, detection segmentation tasks, reveal merits proposed approach, outperforming state-of-the-art methods comparatively lower compute requirements. EdgeNeXt model 1.3M parameters achieves 71.2% top-1 accuracy ImageNet-1K, MobileViT an absolute gain 2.2% 28% reduction FLOPs. Further, our 5.6M 79.4% ImageNet-1K. The code available at https://t.ly/_Vu9 .