An efficient and low power deep learning framework for image recognition on mobile devices

Image classification on mobile devices can provide convenient and secure services for users when using various social software. The traditional classification method mainly relies on the user’s manual marking, but the accuracy of automatic classification has some defects. With the development of convolutional neural network(CNN), the design of lightweight neural network has become a hot topic. However, the state-of-the-art studies always sacrifice classification accuracy for network lightweight, which greatly frustrates usability. In this paper, a new neural network framework, named MobVi, is proposed to enhance the precision of lightweight neural network by solution space division. MobVi is including image solution space division and judgment class. The former uses clustering method based on deep learning to distinguish which small solution space the image belongs to, while the latter uses lightweight neural network customized for the solution space to judge the class. In order to reduce the amount of model parameters and calculations, we designed a customized CNN module. Finally, we propose an energy prediction model to measure whether the model can be successfully implemented on mobile devices. A series of experiments have proved that MobVi has better performance than most existing models for mobile devices. Our model achieves 83.5% accuracy on CIFAR-10 data set, and the parameter quantity is only 2.0 M.