ABSTRACT
This paper details a method for automatic object SAR object identification based on combining the use of residual inaccuracy convolutional neural networks and magnitude reduction of the PCA. The entire method must collect SAR destination picture data and annotation it with item classes to establish a training set for constructing the training set training samples must be enhanced, multiply and pre-processed before utilizing convolutional neural networks with residual inaccuracy. The network model formed is used to transmit training samples into the network for training. When new training sets are premade, the feature vectors are extracted via the network at this moment, it uses magnitude downsized via the PCA method, trained utilizing a multi-class SVM classifier. Finally, by feeding pre-processed sample data into the network for identification and an SVM classifier for identification after PCA magnitude reduction can be used for training SVM. This technique effectively resolves the regular poor identification accuracy difficulty in current SAR Object Identification systems.
