A Research On Convolutional Neural Network Based Domain Adaptation

Domain Adaptation is an important research field to improve the robustness of classifier and strength its practical ability.Deep learning methods are outstanding in extracting features.In order to solve the feature extracting problem and improve the accuracy,our work focused on applying convolution neural network into domain adaptation.This paper researched the problem in three aspects,and proposed three algorithms in supervised,semi-supervised and unsupervised learning correspondingly.The concrete works can be summarized as follows:i)We proposed a supervised learning domain adaptation algorithm by combing CORAL and dCNN.Basing on previous dCNN net,we firstly simplified the net of dCNN,and then input them into correlation alignment(CORAL)algorithm to narrow the variance between source and target domains again,and the results showed that this step can improve the accuracy.ii)We proposed a semi-supervised learning domain adaptation algorithm by training target dataset batch-by-batch.This method based on the assumption that different samples in target domain have different difficulties to classify,so we proposed that we can take samples which have pretty high confidence values as supervised samples and add them into training dataset to reduce the difference between source and target domains.This simple idea has gained pretty high improvements.iii)We proposed a generalize transfer learning clustering model and an unsupervised domain adaptation algorithm based on local kernel learning method.We listed the popular feature extracting and clustering methods,and then put forward a three-stepmodel for cluster problem.Then we combined a simple while novel key-points choosing algorithm into local kernel alignment algorithm which was proposed basing on multikernel clustering algorithm to reduce the computing cost of original algorithm.we extracted the features by pre-trained VGG-net,and then input the features into above-mentioned clustering algorithm,the results showed an satisfying performance than original algorithm.