Evaluating CNNs for Military Target Recognition

Convolutional neural network (CNN) is an efficient algorithm in deep learning. Aiming at the field of military target recognition, this paper constructs a dataset for military target recognition in battlefield, which contains ten kinds of targets. The characteristics of the dataset are described and analyzed. Three classical CNN models (AlexNet, VGGNet and ResNet) and two learning strategies (dropout and data augmentation) are evaluated on the dataset. Under the same condition of the dataset and the same super-parameter setting, the effects of different models are presented and analyzed. The experimental results show that the mean average precisions of ResNet and VGGNet are better than AlexNet, and the accuracies of both ResNet and VGGNet are over 90% with only thousands of training images. At the same time, the dropout and data augmentation strategies have a strong effect for improving the performance.