The thesis deals with the use of super-resolution to improve the quality of lowresolution
images. Super-resolution represents an image processing task that aims
to transform a low-resolution image into a high-resolution image. In the thesis, we
addressed the problem of low quality of input images, which can negatively impact
the performance of various downstream computer vision tasks and applications.
The main goal of the thesis was to improve the quality of these images using
advanced super-resolution techniques such as Real-ESRGAN and ResShift.
The methodology of the work included an overview of the theoretical foundations
of super-resolution, development and implementation of Real-ESRGAN
and ResShift models, and their testing on low-resolution images. Various quality
metrics such as PSNR, SSIM, BRISQUE, and NIQE were used to objectively
evaluate the capabilities of the tested superresolution techniques.
The results of our experiments showed that the used models can significantly
improve the quality of low-resolution images. More specifically, the ResShift
model generally performs better on the PSNR and SSIM metrics compared to
the Real-ESRGAN model, that it can ensure better reconstruction quality.
|
