TripleNet: A Low-Parameter Network for Low Computing Power Platform

Deep learning has achieved great success in computer vision (CV), and convolutional neural networks (CNN) have become an important network architecture for CV tasks. With the widespread use of mobile devices, neural network models based on low-computing power platforms have received much attention. This paper proposes a lightweight CNN model named TripleNet. TripleNet learns the model compression methods of HarDNet and ThreshNet and improves on them. TripleNet completes the network architecture by combining three different convolutional layers, which has the advantages of small model size and high test accuracy. With similar test accuracy, TripleNet model has fewer parameters than HarDNet and ThreshNet. This paper conducts image classification experiments using TripleNet and other networks on CIFAR-10 and SVHN datasets. The experimental results show that compared with HarDNet, the parameters of TripleNet are reduced by 66% and the accuracy rate is improved by 18%; compared with ThreshNet, the parameters of TripleNet are reduced by 37% and the accuracy rate is improved by 5%.