A model integrating attention mechanism and generative adversarial network for image style transfer

Abstract

Image style transfer is an important way to combine different styles and contents to generate new images, which plays an important role in computer vision tasks such as image reconstruction and image texture synthesis. In style transfer tasks, there are often long-distance dependencies between pixels of different styles and contents, and existing neural network-based work cannot handle this problem well. This paper constructs a generation model for style transfer based on the cycle-consistent network and the attention mechanism. The forward and backward learning process of the cycle-consistent mechanism could make the network complete the mismatch conversion between the input and output of the image. The attention mechanism enhances the model’s ability to perceive the long-distance dependencies between pixels in process of learning feature representation from the target content and the target styles, and at the same time suppresses the style feature information of the non-target area. Finally, a large number of experiments were carried out in the monet2photo dataset, and the results show that the misjudgment rate of Amazon Mechanical Turk (AMT) perceptual studies achieves 45%, which verified that the cycle-consistent network model with attention mechanism has certain advantages in image style transfer.

Keywords

• Image style transfer
• Cycle-consistent network
• Attention mechanism
• Long-distance dependencies