Research On Cross-Domain Image Recognition Algorithm Based On Pair-Wise Generalization Network And Attention Mechanism

In recent years,deep learning networks have become a popular choice in machine learning and computer vision task.Although there have been many deep learning network algorithms,these methods often assume that the data come from the same domain.When these methods are applied to different domains,their effects will decrease significantly.However,in the real world,the samples of some domain are sometimes unavailable or expensive to label it,so it is difficult to be learned by the traditional deep learning algorithm.The cross-domain learning method came into being.The cross-domain learning is also known as transfer learning,it solves these problem by applying knowledge or patterns learned in a particular domain to a different but related domain.Based on the development trend of deep learning network,there are mainly two parts in the research work of this paper:(1)Cross-domain image recognition algorithm based on Pairwise Generalization Network;(2)Cross-domain image recognition algorithm based on Pairwise Attention Network.1.This thesis proposes a cross-domain image recognition algorithm based on Pairwise Generalization Network.It adds Instance Normalization and Batch Normalization to enable CNN models for enhancing their performance in the original domain and transfer to the new domain.We use the Siamese architecture to learn a discriminative embedding subspace,and make positive sample pairs aligned and negative sample pairs separated.This can work well even with only few labeled target data samples.We also add residual architecture and MMD loss for our model to improve the performance.Extensive experiments on two different public benchmarks,MNIST-USPS and Office-31 show that the cross-domain image recognition algorithm based on Pairwise Generalization Network significantly outperforms the state-of-the-art methods.Compared with CCSA algorithm,this algorithm is 4.9%better in MNIST-USPS and Office-31 experiments.2.This thesis proposes a cross-domain image recognition algorithm based on Pairwise Attention Network.It connects local-feature and global-feature into an attention map to focus on important parts of the image.We also use the Siamese network to learn a discriminative embedding subspace that is but different with majority method,it maps positive sample pairs aligned in a hypersphere and negative sample pairs separated.We also add an attention consistency as part of the learningprocess to make sure consistent interest regions in the same class.We conduct extensive evaluations on the MNIST-USPS,Office-31 and Visda-2017 dataset,and the cross-domain image recognition algorithm based on Pairwise Attention Network has surpassed the state-of-the-art methods in these datasets.