Yiwen Xu, Tariq M. Khan, Yang Song, Erik H. W. Meijering
A comprehensive survey paper summarizing the foundational principles, hardware classification, model lightweighting techniques, and applications of edge deep learning in computer vision and medical diagnostics.
There is a lack of systematic guidance and up-to-date trends for efficiently deploying deep learning models on edge devices. In particular, specific methodologies and hardware selection criteria are needed for real-time applications such as medical diagnostics.
Introduce the basic concepts and advantages of edge deep learning, and propose a new classification system for edge hardware platforms based on performance and usage scenarios. Additionally, review in detail lightweight network design and model compression techniques (pruning, quantization, distillation, etc.), and analyze practical application cases in general computer vision and medical diagnostics.
Provide a comprehensive reference for edge deep learning researchers and practitioners. Systematize hardware classification and model optimization methodologies to offer practical guidelines for deploying deep learning in edge environments. Future research directions and challenges are also discussed.