Accelerating Convolutional Neural Networks for Continuous Mobile Vision via Cache Reuse

Convolutional Neural Network (CNN) is the state-ofthe-art algorithm of many mobile vision fields. It is also applied in many vision tasks such as face detection and augmented reality on mobile devices. Though benefited from the high accuracy achieved via deep CNN models, nowadays commercial mobile devices are often short in processing capacity and battery to continuously carry out such CNN-driven vision applications. In this paper, we propose a transparent caching mechanism, named CNNCache, that can substantially accelerate CNN-driven mobile continuous vision tasks without any efforts from app developers. To cache and reuse the computations of the similar image regions which are consecutively captured by mobile devices, CNNCache leverages two novel techniques: an image matching algorithm that quickly identifies similar image regions between images, and a cache-aware CNN inference engine that propagates the reusable regions through varied layers and reuses the computation results at layer granularity. We implement a prototype of CNNCache to run on commodity Android devices, and evaluate it via typical CNN models. The results show that CNNCache can accelerate the execution of CNN models by 20.2% on average and up to 47.1% under certain scenarios, with no more than 3.51% accuracy loss.