Design of High Performance Convolutional Neural Network for Lightweight Platform

Abstract

Machine learning plays an important role in the field of data science. With the further promotion of deep learning, it is becoming more and more valuable to study deep learning network models that can efficiently run in distributed computing, cloud services and mobile platforms. However, due to the limited bandwidth, storage and operation performance, it is still a huge challenge to accurately and efficiently complete neural network computing on these platforms. In this paper, a lightweight convolution neural network called FccNet(fusion cross convolution network) is proposed. This new convolution neural network architecture extracts multi-scale features by small-scale convolution, and combines the characteristics of parallel crossover, identity mapping and feature fusion. In order to evaluate the performance of FccNet comprehensively, this paper conducted experiments on the Food-101, Caltech-256 and GTSRB Image Classification Benchmark datasets. FccNet can significantly reduce the model parameters and greatly improve the performance.