Facial Attribute Editing based on Independent Selective Transfer Unit and Self-attention Mechanism

Facial attribute editing aims to change the facial attributes, which can be regarded as an image translation problem. Facial attribute editing is usually realized by combining encoder-decoder and Generative Adversarial Networks, but the generated image is not realistic enough, and the model has weak ability to control the fine granularity of face attributes of generated images. In this work, we propose a Generative Adversarial Network ISTSA-GAN based on Independent Selective Transfer Unit (ISTU) and Self-attention Mechanism. On the basis of STGAN, we use ISTU instead of Selective Transfer Unit (STU) to combine with encoder-decoder to selectively transfer the features of encoder. In addition, a self-attention mechanism is introduced into the transposed convolution layer of the decoder to establish long-distance dependence of the model across image regions. Finally, attribute interpolation loss and source domain adversarial loss are added to constrain the training of the model. Experimental results show that this method can improve the ability of editing attributes and saving much details, and enhance the ability of fine-grained control of editing attributes. It is superior to classical methods in attribute editing accuracy and image quality.