StereoSqueezeNet: With fewer parameters but higher accuracy than SqueezeNet

Abstract

Convolutional neural networks (CNNs) have evolved from the initial LeNet to date, and network models have become increasingly deep and comprehensive. It has been proven that deeper networks have better fitting effects, but the corresponding parameter size and computational complexity increase rapidly. With the continuous development of mobile Internet technology, portable devices have been rapidly popularized, and users have put forward more and more demands. Thus, how to design efficient and high-performance lightweight convolutional neural networks (CNNs) is the key to solve this challenging problem. Recently, this type of convolutional neural networks (CNNs)--lightweight convolutional neural networks (CNNs), which adopt the design concept of compression networks and maintain high accuracy with fewer parameters, has attracted increasing attention. SqueezeNet is a lightweight CNN adapting to edge device deployment. Its number of parameters is only 1/50 of AlexNet, but it achieves the same accuracy as AlexNet. In order to make the network more lightweight, inspired by SqueezeNet, MobileNet, SENet, SKNet, AlexNet, etc., in this paper we propose StereoSqueezeNet, using much fewer parameters but achieving even better accuracy than SqueezeNet.