An AI-based image enhancement system with its FPGA implementation

• Main Author: Yang, Hao
• Other Authors: Gwee Bah Hwee
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Electrical and electronic engineering::Integrated circuits
• Online Access: https://hdl.handle.net/10356/168449
• Institution: Nanyang Technological University

One of the essential parts of image processing is image enhancement. With the contribution of artificial intelligence (AI), this dissertation proposes a novel deep learning system for image enhancement. The proposed network is based on the structure of U-Net, and it is capable of image enhancing and denoising simultaneously. The experiment results of this system show a significant performance improvement compared to conventional systems in adaptive methods. By introducing pixel shuffle algorithms from super-resolution, we eliminate checkerboard artifacts significantly. Finally, the proposed network achieves a quantitative evaluation with PSNR/SSIM is 20/0.85 with the post-trained model. This proposed system could be implemented with the Xilinx FPGA platform. Furthermore, an FPGA platform that runs NVDLA as a hardware backend has been implemented and tested with Lenet5 and Resnet18.