Optimization of Convolutional Neural Network Target Recognition Algorithm

This paper proposes an optimized convolutional neural network target recognition algorithm for the problem of low recognition rate of synthetic aperture radar (SAR) target training, under the condition of insufficient tag data, translation, rotation and complexity. In order to overcome the shortage of tag data, the convolutional neural network is initialized with a feature set, obtained by principal component analysis (PCA) unsupervised training. In order to improve the training speed while avoiding overfitting, Rectified Linear Unit (ReLU) function is used as the activation function. In order to enhance robustness and reduce the effect of down sampling on feature representation, this work uses a maximum probability sampling method and normalizes the local contrast of feature after convolution layers. The experimental result shows that, compared with traditional convolutional neural network, this approach achieves a higher recognition rate for SAR target and better robustness to various image deformation and complex background.