A Muti-branch Convolutional Netural Network for Mobile Platform

Abstract

We proposed a new type of light-weight convolutional neural network MRINet for low computing power requirements. This model is applied with strategies including depthwise separable convolution, Channel pruning and ELU activation. It greatly reduces the amount of calculation while getting high accuracy. MRINet can complete the training process and application on mobile phones and other mobile platforms. By integrating into the corresponding application, which can solve many real problems including self-medication. By training on the ISIC dataset, as compared to MobileNet, we improved training speed by 23%, while accuracy is increased 3.3%.