COMPARISON OF COLOUR CONSISTANCY ALGORITHMS FOR DIGITAL IMAGES

V diplomskem delu raziskujemo barvno vtzrajnost za digitalne slike. V uvodu predstavimo problem osvetlitve pri analizi barvnih slik. Barvno vztrajnost povezujemo s prilagajanjem osvetlitvi, ki ga zmore človeški in živalski vid. Naš cilj je preučiti obstoječe algoritme za barvno vztrajnost in nato razviti lasten algoritem, ki bi ga lahko uporabili pri nadaljnjem raziskovanju na področju računalniškega vida. Najprej smo se seznanili s programskimi orodji, ki že vključujejo pogosto uporabljane algoritme za obdelavo slik in video posnetkov. Podrobneje opisujemo tudi rutine in koncepte, ki smo jih srečali pri poglobljenem študiju algoritmov za barvno vztrajnost. Nekatere od njih smo izbrali kot izhodišče pri razvoju lastne metode za barvno korekcijo slik. Dodatno poglavje namenjamo barvnemu kalibratorju. Uporabljamo ga za ocenjevanje algoritmov in tudi kot referenco pri nadzorovanem strojnem učenju. Osrednji del diplome je posvečen razvoju in opisu lastne metode. Razvili smo jo kot obteženo vsoto obstoječih algoritmov, ki imajo komplementarne lastnosti, tako da njihova združitev odpravlja slabosti posameznih pristopov. Uteži za linearne kombinacije petih izbranih algoritmov smo določali s štirimi metodami, in sicer od preproste z enakimi utežmi do najzahtevnejše s strojnim učenjem. V eksperimentih smo za vrednotenje rezultatov in delovanja metod uporabili barvni kalibrator. Ugotavljamo, da postopek za barvno vztrajnost, ki smo ga razvili, dosega boljše rezultate kot posamezni algoritmi, iz katerih je sestavljen.We study colour constancy for digital images in this diploma thesis. The illumination problem when analyzing colour images is presented in introduction section. Colour constancy is associated with the illumination adaptation that is capable human and animal sight. Our aim is to study existing colour constancy algorithms and develop our own algorithm that we could use for further research on the computer vision field. First, we familiarize ourselves with programming tools that implement frequently used image and video processing algorithms. We describe in detail methods and concepts that we have encountered by in-depth study of colour constancy algorithms. Some of these algorithms are chosen as starting point for developing our own method for colour correction of images. A special chapter is dedicated to the colour checker card. This card is used for algorithm evaluation and as the ground truth for supervised machine learning. The central part of our thesis is dedicated to a development and description of our own method. We develop this method as a weighted sum of existing algorithms that have complementary properties, so that their combination eliminates their individual weaknesses. Weights for the linear combinations of five selected algorithms are assigned by using four methods, namely from a simple with uniform weights to the most complicated by using machine learning. A colour checker card is used for evaluating the results in our experiments. We ascertain that the colour constancy method, we have developed, achieves better results than individual algorithms, from which it is composed