A lightweight convolutional neural network model for target recognition

Abstract

Abstract Convolutional neural networks have achieved excellent performance in a wide range of applications, but the huge resource consumption makes a great challenge to their application on mobile terminals and embedded devices. In order to solve such problems, it is necessary to balance the size, speed and accuracy of the network model. This study proposed a new shallow neural network on the bases of ResNet and DenseNet. We use different size convolution kernels to obtain feature maps and then concat them. Afterwards we build two convolution layers to reduce the size of the feature maps and increase the depth of the network. By stacking this structure, we get our net model. Experiments show that our nine-layers network recognition performance is better than 18-layers ResNet and 19-layers DenseNet, and its training time is shorter. The final recognition rate of our network is 97.37%, ResNet recognition rate is 96.93%, and DenseNet is 96.31%.