Research On Deep Visual Domain Adaptation Based On Adversarial Learning

Recent years,with the development of deep learning and convolutional neural networks(CNN),the ability of algorithm in visual tasks(image classification,object detection and semantic segmentation,etc.)has been greatly improved.Deep models always rely on a large number of labeled training data for achieving a good performance.However,labeling massive data is a time-consuming and laborious process.To avoid high-cost labeling process and make effective usage of accessible unlabeled test data,domain adaptation can transfer the learned knowledge from training data to test data for improving the performance.It aligns the distribution of training data domain(source domain)and test data domain(target domain).In this paper,the adversarial learning technique in Generative Adversarial Networks is used for unsupervised domain adaptation.Through adversarial training,the generative network can be trained to transform the source image to a new image similar with target image without any labels in target domain.Finally,the target classifier is trained with the generated image and the label corresponding to the source domain.Then a model with good performance on the test data can be obtained.Based on the above framework,two improved algorithms are proposed in this paper:We propose an attention-based adversarial domain adaptation algorithm.In order to keep the semantics of the transformed image consistent with the source image,a semantic consistency constraint is defined without requirement of additional labels or mask information.The proposed constraint constrains the global and local semantics of the image.The generated image not only has the same prediction label as the source image,but also has the same attention information coded by convolutional neural network as the original image,which is called attention alignment,inspired by the visual attention mechanism in deep learning.The proposed attention constraint in spatial level improves the quality and structure stability of the generated images.We propose Wasserstein distance guided adversarial domain adaptation algorithm,in which the traditional objective function is replaced by the adversarial loss based on Wasserstein distance.The method proposed in this paper directly uses the data distribution of source and target domains to estimate empirical Wasserstein distance,rather than feature distribution.This method can estimate the difference between source and target domains more accurately.The proposed algorithm can improve the stability of adversarial learning and effectively alleviate the phenomenon of mode collapse.Finally,the improved algorithm is evaluated on SVHN-MNIST and Office-Home Product-Clipart domain adaptation tasks.The experimental results show that the two improved algorithms both improve the quality of generated samples and the classification accuracy of the target classifier and they are superior to the existing domain adaptation algorithms,such as transfer component analysis,domain adversarial neural networks,etc.At the same time,the two algorithms have their own advantages and disadvantages.The Wasserstein distance guided adversarial domain adaptation algorithm is faster,while the attention-based adversarial domain adaptation algorithm can handle more complex tasks and is more general.