Implementation and comparison of super resolutionalgorithms on embedded systems

Abstract

In this thesis, we aim to implement CNN based super-resolution methods and our implementation is image and video on embedded systems. CNN based super-resolution methods are EDSR, ESPCN, FSRCNN and LAPSRN. We compared super-resolution methods based on scale factor, layers and parameters. We trained all these super-resolution methods on embedded systems. These methods are explained and applied on embedded systems. We compared interpolation-based methods, reconstruction methods and learning-based methods. We focused on performance enhancing research to achieve real-time performance and we implemented the super-resolution implementation in 2 processes. They are training process and implementation process. In training process, we used different datasets on all super-resolution methods and we produced trained files, which have all scale factor for implementation process. First of all in implementation process, we focused on choosing the programming language for better performance. We chose a super-resolution method, we produced same code on different programming languages for super-resolution implementation and we analyzed implemetation performance on different programming languages, then we analyzed PSNR, SSIM and elapsed time values on each CNN based super-resolution methods for quality. We found a trade off for quality and processing time. To achieve real-time performance, we implemented to image and video implementations on CPU and DPU and we compared the performance of super-resolution implementations on CPU and DPU. We obtained to inferences to reach real-time super-resolution implementation. We used single-core and multi-core structure on image implementation and we compared single-core and multi-core CPU performance on image implementation with using all super-resolution methods. We used a frame in single-threading and a frame which consisting of many tiles in multi-threading structure on video implementation. We created a image by stitch all the tiles on video implementation. CPU and DPU analyzes are reported for real time performance. We shared the codes of super-resolution applications and analysis on our Github account.