Application of Convolutional Models for Image Superresolution

Uvijek je poželjno imati sliku velike rezolucije. Od manje slike moguće je tradicionalnim metodama kao što su bilinearna i bikubična interpolacija povećati sliku, ali dolazi do uvođenja nazubljenih rubova, zamućenosti i povećanja šuma. U novine se vrijeme predlažu modeli zasnovani na dubokom učenja koji povećavaju rezoluciju slika pritom rekonstruirajući visokofrekventne detalje. Ti modeli primjenu nalaze na području medicinske i satelitske fotografije te za opću primjenu u medijskim sustavima. U ovom radu implementirana su tri modela: WDSR, DRLSR i ESPCN. Korišteni su PSNR i SSIM kao ocjene kvalitete. Implementirani modeli su učeni na DIV2K skupu, a testirani na skupovima Set5, Set14, T91, Urban100, BSDS100, BSDS200, Manga109, General100, Historical, Aerial, PROBA-V i DRIVE.Having a large resolution image is always desired. It is possible to transform a low resolution image to a high resolution image using the classical methods such as bilinear and bicubic interpolation. However they also introduce jaggedness, blurriness and noise amplification. Recently deep learning models have been proposed for generating images with high frequency details. These models are used in medical and satelite imagery as well as general media. In the context of this thesis three models have been implemented: WDSR, DRLSR and ESPCN. PSNR and SSIM are used as measures of quality. The models have been trained on the DIV2K dataset and tested on Set5, Set14, T91, Urban100, BSDS100, BSDS200, Manga109, General100, Historical, Aerial, PROBA-V and DRIVE datasets