Transfer Learning for Mobile-Based Computer Vision

Recent years have witnessed rapid growth in mobile-based computer vision applications.Deep learning advances have led to this, but the models have intensive computational and memory requirements that make their deployment on resourceconstrained mobile devices challenging.Transfer learning offers a solution through the reuse of pre-trained models that reduce training time and data requirements.This paper explores how transfer learning has been applied for mobile-based computer vision, with particular emphasis on model development acceleration that is achieved by preserving performance.In particular, we discuss strategies to leverage a pre-trained model in overcoming some of the restrictions imposed by hardware on mobile.In addition to that, the paper highlights several best practices toward optimizing pre-trained models for application on mobile systems, including compressing models as well as how real-time performances can be optimized.We show through case studies and experimental results how transfer learning can balance efficiency with accuracy, making it an ideal approach for deploying computer vision models on mobile devices.