MobiVision: A Novel Energy-Efficient Mobile Deep Learning Framework for Computer Vision

摘要

The development of mobile devices, such as smartphones, drones and augmented-reality headsets, have greatly promoted processing the convolutional neural network (CNN) model on them. One popular research topic is designing CNN models for image classification task on mobile terminals because these equipments would produce many videos or images data everyday generally. However, these pre-trained models usually possess complex structure and plenty of parameters so that they are difficult to be implemented on resource-limited mobile terminals for their serious time delay and energy consumption. A common solution is compressing neural networks to make them adapt to limited computation and memory resource in mobile devices, but it is not the best idea for pruned models always sacrificing accuracy. In this paper, We propose MobiVision, a novel neural network framework that conclude two main stages, which is defined as partitioning solution space and judging class for an image input. The former, utilizing deep learning-based clustering method, focuses on distinguishing which small solution space an image belongs to, while the latter calls a light-weight neural network associated to that solution space to recognize certain class of input. Series of experiments have proved that MobiVision achieves better performance than most of existing models serving for mobile devices because energy MobiVision consumed is little as well as accuracy of the model is equivalent to others meanwhile.