Real-time retinex-based and learning-based methods for computational color constancy with high accuracy

doctoral thesis

• Title (croatian): Metode visoke točnosti za računalnu postojanost boja za rad u stvarnom vremenu temeljene na Retinex pristupu te učenju
• Author: Nikola Banić
• Mentor: Sven Lončarić (mentor)
• Committee member: Sven Lončarić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; (Department of Electronic Systems and Information Processing); Zagreb
• Defense date and country: 2016, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing; Data Processing
• Universal decimal classification (UDC): 621.3 - Electrical engineering

Abstract

Removing the influence of illumination and camera characteristics on image colors is an important image enhancement problem. It is solved by applying computational color constancy methods, whose goal is to assure that colors of objects in the image remain relatively constant regardless of the illumination conditions in the scene. Many such methods have been proposed and they either try to mimic the human visual system's color constancy or simply to produce visually appealing results by various means. These methods can operate locally by changing the color of each pixel individually or globally by changing all pixels in the same way. When the image scene is lit by several illumination sources, the application of local instead of global methods provides better results. However, the computational cost of local methods is higher than for global methods. The global methods are best applied when there is uniform illumination, their accuracy is relatively easily measured and evaluated, and they are faster than local methods. The latter is the main reason why they are also implemented in most digital cameras. Although many local and global color constancy methods have been proposed, neither of them gives satisfactory results in all conditions. On the other hand, some of the methods that give better results than other methods are not always suitable for real-time applications. All this means that there is still much space for improvements in terms of accuracy, resulting image quality, and computational speed. Guided by this need for these improvements, in this thesis several local and global color constancy methods suited for cases where high computational speed or high accuracy is needed are proposed. These methods can be summarized as follows: • a fast local method for color constancy and brightness adjustment in non-uniform illumination based on the Retinex theory of color vision; • a fast global Retinex-based method for color constancy in uniform illumination; and • a highly accurate global learning-based method for color constancy in uniform illumination using simplified models of possible illumination values. Experimental results of all of the proposed methods are presented and discussed. The proposed local methods are shown to produce results of high quality faster then some other methods used for the same purpose. As for the proposed global methods, in most cases the learning-based ones outperform the majority of other state-of-the-art methods in terms of accuracy. The proposed low-level statistics-based methods outperform all other methods in terms of computational speed while also having a relatively high accuracy and being hardware-friendly.

Abstract (croatian)

Uklanjanje utjecaja osvjetljenja i kamere na karakteristike boja slike je važan problem poboljšanja slike. Rješava ga se primjenom računalnih metoda postojanosti boje čiji je cilj osigurati da boje objekata u slici ostanu relativno postojane neovisno o uvjetima osvjetljenja u sceni. Predloženo je mnogo takvih metoda i one ili oponašaju postojanost boja ljudskog vizualnog sustava ili jednostavno na različite način daju vizualno privlačne rezultate. Te metode mogu djelovati lokalno mijenjajući boju svakog piksela posebno ili globalno mijenjajući sve piksele na jednak način. Kad scenu slike osvjetljava više izvora svjetlosti, primjena lokalnih umjesto globalnih metoda daje bolje rezultate. Ipak, lokalne metode imaju veću računsku cijenu nego globalne. Globalne metode je najbolje primijeniti pri jednolikom osvjetljenju, točnost im je lako mjeriti i vrednovati i brže su od lokalnih metoda. Iako je predloženo mnogo lokalnih i globalnih metoda za postojanost boja, nijedna od njih ne daje zadovoljavajuće rezultate u svim uvjetima. S druge strane neke od metoda koje daju bolje rezultate nisu uvijek prikladne za primjene u stvarnom vremenu. Sve to znači da još uvijek ima prostora za poboljšanje točnosti, kvalitete rezultantne slike i računske brzine. Pod idejom potrebe za tim poboljšanjima u ovoj disertaciji predloženo je nekoliko lokalnih i globalnih metoda za postojanost boja koje su prilagođene za slučajeve gdje su potrebni visoka računska brzina ili visoka točnost. Te metode su sljedeće: • brza lokalna metoda za postojanost boja i podešavanje svjetline kod nejednolikog osvjetljenja temeljena na Retinex-teoriji vida u boji; • brza globalna metoda temeljena na Retinexu za jednoliko osvjetljenje; i • globalna metoda za postojanost boja visoke točnosti temeljena na učenju za jednoliko osvjetljenje koja koristi pojednostavljene modele mogućih vrijednosti osvjetljenja. Prikazani su i raspravljeni eksperimentalni rezultati predloženih metoda. Pokazano je da predložene metode daju rezultate visoke kvalitete brže nego neke druge metode koje se koriste za istu stvar. što se tiče globalnih metoda, u većini slučajeva one temeljene na učenju su točnije od većine drugih vrhunskih metoda. Predložene statističke globalne metode su brže od svih drugih metoda pri čemu također imaju relativno visoku točnost i pogodne su za sklopovsku implementaciju.

• Keywords: color constancy; illumination estimation; image enhancement; Retinex; white balancing.
• Keywords (croatian): postojanost boja; procjena osvjetljenja; poboljšanje slike; retinex; podešavanje bijele.
• Language: english
• URN:NBN: urn:nbn:hr:168:307486
• Study programme: Title: Electrical Engineering and Computing; Study programme type: university; Study level: postgraduate; Academic / professional title: Doktor znanosti elektrotehnike i računarstva (Doktor znanosti elektrotehnike i računarstva)
• Type of resource: Text
• Extent: 102 str. ; 30 cm.
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2019-04-11 13:01:46