Evaluation of neural network based image super-resolution
1University of Oulu, Faculty of Information Technology and Electrical Engineering, Computer Science
|Online Access:||PDF Full Text (PDF, 6.9 MB)|
|Persistent link:|| http://urn.fi/URN:NBN:fi:oulu-202212143760
Oulu : N. Salimi,
|Publish Date:|| 2022-12-14
|Thesis type:||Master's thesis (tech)
Super-resolution (SR) aims to produce a higher resolution image containing more details than the original image. The amount of pixels is easy to add with simple interpolation methods, but the amount of details does not increase. To overcome this limitation single image super-resolution (SISR) was introduced, which aims to recover the high-resolution (HR) image from the low-resolution (LR) images.
Convolutional neural networks (CNN) have become an essential method in machine learning. With the growth of CNN, super-resolution solutions have grown immensely. In this work, a broad review is done on neural network methods designed for super-resolution. Four methods are chosen by their originality and different architectural choices, implemented in PyTorch framework. The models are already trained with public datasets, and the pre-trained models are used for the evaluation. The evaluation is done by analyzing the results with qualitative and quantitative methods.
All the methods are tested with public datasets and a private dataset called Hiottu-1, including a wood surface images with different defect types. The evaluation is done based on their image quality and inference time. Peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) metrics are used for quality evaluation, and the inference time is measured by how fast the model generates the output result of test image.
The chosen methods improved the image qualities of test images in each datasets. The best perfoming ones were swin image restoration (SwinIR) and pixel attention network (PAN) methods. SwinIR had better PSNR and SSIM values than PAN method and results were pealing to human eye. The inference time of SwinIR is slow, therefore the best possible application would be offline usage. The PAN method had impressing results and its inference time enables the real-time application usage.
The SwinIR performed extremely well on Hiottu-1 dataset, with increasing the image quality of defect types and reducing noise overall. The PAN method got high metrics values on Hiottu-1 dataset, although the results were not as pealing as the SwinIR. In the wood manufacturing inspection side, the SwinIR could be utilized on slow production line with high defect detection accuracy, while the PAN method could be utilized on faster production line.
© Nechir Salimi, 2022. This publication is copyrighted. You may download, display and print it for your own personal use. Commercial use is prohibited.