Cross-Domain Variational Adversarial Autoencoder

The existing cross domain image generation algorithms usually require users to provide pairwise data,and the generation ability is limited,which only supports one-to-one cross domain image generation.To solve the above problems,a cross-domain variational adversarial autoencode framework is proposed,which realizes one-to-many continuous transformation of cross-domain images without providing any paired data.Assuming that images from different domains share the same content attributes and have independent style attributes,the one-to-many continuous transformation of cross-domain images can be converted into four steps,including decoupling,encoding,fitting and cross-domain stitching of content attributes and style attributes.Firstly,a content code and a style code of the cross-domain data are decoupled using the proposed encoder.Then the two kinds of codes are fitted by the adversarial operation and the variational operation respectively.Finally,the image reconstruction is realized by stitching the content code and the style code of the single-domain image,and the one-to-many cross-domain continuous transformation is obtained by cross-stitching the content code and the style code of the different-domain images.The supervised cross domain image generation results on the standard datasets MNIST and SVHN satisfy both the authenticity and diversity,and are superior to other cross domain image generation algorithms in the quantitative evaluation of classification accuracy and domain self-adaptability.In addition,the unsupervised cross domain image one-to-one generation is realized on the face dataset NIR-VIS and the sketch dataset Edges-Shoes,which visualization results present that the feature distribution of the generated image is consistent with the feature distribution of the source images.The above experiments demonstrates the feasibility and effectiveness of the proposed framework.