Caffe2C: A Framework for Easy Implementation of CNN-based Mobile Applications

Abstract

In this study, we create "Caffe2C" which converts CNN (Convolutional Neural Network) models trained with the existing CNN framework, Caffe, C-language source codes for mobile devices. Since Caffe2C generates a single C code which includes everything needed to execute the trained CNN, csCaffe2C makes it easy to run CNN-based applications on any kinds of mobile devices and embedding devices without GPUs. Moreover, Caffe2C achieves faster execution speed compared to the existing Caffe for iOS/Android and the OpenCV iOS/Android DNN class. The reasons are as follows: (1) directly converting of trained CNN models to C codes, (2) efficient use of NEON/BLAS with multithreading, and (3) performing pre-computation as much as possible in the computation of CNNs. In addition, in this paper, we demonstrate the availability of Caffe2C by showing four kinds of CNN-base object recognition mobile applications.