Convolutional neural networks recognition algorithm based on PCA

To improve the insufficiency of Synthetic Aperture Radar(SAR) labeled training data for Convolutional Neural Networks(CNN) and the recognition rate for large variations, a novel CNN recognition algorithm is proposed. Firstly, a set of features is extracted from the original data by unsupervised training based on PCA as the initial filter set for CNN. Secondly, in order to accelerate the training speed while avoiding over-fitting, the Rectified Linear Units(ReLU) is adopted as the non-linear function. Thirdly, to strengthen robustness and mitigate the defects of pooling upon features, a probabilistic max-pooling sampling method is introduced and local contrast normalization is exploited on features after the convolutional layer. Experiments demonstrate that our algorithm outperforms the original CNN in recognition rate and achieves better robustness for large variations and complex background.