Superločljivost z aplikacijo na slikah uhljev

Super-resolution (SR) is a class of image enhancing methods that boosts the resolution of the image. This is useful in various areas, such as visually enhancing photographs or improving person recognition performance. This undergraduate thesis focuses on Single Image Super Resolution of ears, a method of super-resolution that creates missing information from a single image. One of the earliest ways to address the issue of super-resolution was interpolation, but achieved limited success. The latest improvements in SR that have been made feasible by deep neural networks, which significantly improved performance. We evaluated the performance of the Enhanced Deep Residual Network (EDSR) and Shifted Windows Transformer Network (SwinIR) for image super-resolution of ears. Using the AWE dataset which consists of $16,665$ images of ears of various sizes, shapes, and orientations, we trained four models: two on EDSR and two on SwinIR networks, each with scaling factor of two and four. Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) performance measures were used to evaluate the two different model designs. SwinIR achieves a superior PSNR and SSIM, however, the visual results seem to be highly similar.Superločljivost je razred metod za izboljšavo slik, ki povečajo ločljivost slike. To je uporabno na več področjih, kot je npr. vizualna izboljšava fotografij ali izboljšava zmogljivosti prepoznave oseb. Diplomska naloga se osredotoča na superločljivost posamičnih slik uhljev, metodo superločljivosti, ki na posamezni sliki skuša rekonstruirati manj-kajoče dele informacije. Prvi pristopi k superločljivosti so temeljili na interpolaciji, z omejeno uspešnostjo. Globoke nevronske mreže so omogočile naj-novejše doprinose k superločljivosti in pomembno izboljšale delovanje. V okviru naloge smo ocenili delovanje {it Enhanced Deep Residual Network} (EDSR) in {it Shifted Windows Transformer Network} (SwinIR) za slikovno superločljivost uhljev. Z uporabo zbirke podatkov AWE, ki vsebuje 16.665 slik uhljev različnih velikosti, oblik in orientacij, smo naučili štiri modele: dva na EDSR in dva na SwinIR mreži, vsako s faktorjem razširitve dva in štiri. Za ocenitev modelov sta bili uporabljeni metriki najvišjega razmerje med signalom in šumom (PSNR) in merilo indeksa strukturne podobnosti (SSIM). SwinIR doseže boljši PSNR in SSIM kot EDSR, na podlagi vizualnih rezultatov pa sta si metodi močno podobni