Controllable Person Image Synthesis GAN and Its Reconfigurable Energy-efficient Hardware Implementation

Abstract

At this stage, how to controllably generate higher quality person image is still the challenge of person image synthesis. At the same time, the update of image synthesis network is far ahead of its hardware implementation. Therefore, this paper proposes a GAN network for person image synthesis that can generate high quality person image with controllable pose and attributes. The newly designed network is more convenient for hardware implementation while ensuring that the generated image is controllable. This paper also designs a synthesizable library for GAN to pursue faster hardware reconfiguration. We completed the new model proposed in this paper based on this library. Finally, the proposed network achieves better results both quantitatively and qualitatively compared with previous work. Compared with GPU and CPU, the hardware implementation based on FPGA can achieve the highest energy efficient of 73.67 GOPS / W.