Research On Target Recognition Of SAR Images Based On Transfer Learning

When learning a classification model for a new target domain with a small amount of training samples,brute force use of machine learning algorithms often leads to extensive overfitting with poor generalization skills.On the other hand,collecting a sufficiently huge amount of annotated image samples has proven to be rather expensive.Transfer Learning methods aim to solve this kind of problem by transferring knowledge from related source domain which has much more data to aid classification in the target domain.This thesis investigates a complete procedure of the design,implementation and test on how labels learned with deep Convolutional Neural Networks(CNNs)on enormous annotated datasets are transferred effectively to a Synthetic Aperture Radar(SAR)image target recognition task with less amount of training data.This thesis,whose model is based on a deep pre-trained convolutional neural network,puts forward a lightweight feature extractor-based transfer learning method capable of transferring knowledge from VGG-Net as source domain to a SAR target domain with the goal of target recognition of SAR images.While a lean network is adopted to speed up training thereby mitigating potential computational cost,a data process pipeline taking into account background noise suppression is employed to address the data acquisition limitation and reduce redundancy for SAR target recognition.The trained model's general classification accuracy for the integrated data sets averages95%,with near perfect intra-class error-free predictions.Furthermore,our results demonstrate the effectiveness of transfer learning and shows that leveraging knowledge from related categories can improve performance when the training data is scarce.