Variational methods for restoration of images degraded by fog and rain

doctoral thesis

• Title (croatian): Varijacijske metode za obnavljanje slika degradiranih maglom i kišom
• Author: Vedran Stipetić; MBZ: 382192
• Mentor: Sven Lončarić (mentor)
• Committee member: Damir Seršić (predsjednik povjerenstva)
• Committee member: Marko Subašić (član povjerenstva)
• Committee member: Viktor Sučić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; (Department of Electronic Systems and Information Processing); Zagreb
• Defense date and country: 2023-01-27, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing; Data Processing
• Universal decimal classification (UDC): 004 - Computer science and technology. Computing. Data processing

Abstract

The number of images captured every day increases as time goes on. Many of these images are taken outdoors, where weather effects can influence the quality of the captured image. The degradation caused by atmospheric effects can be problematic, both in terms of visibility for a human observer and reducing the efficacy of computer vision algorithms. Because of this, the task of restoring these images is essential. This thesis focuses on restoring images degraded by fog or rain. The current state-of-the-art methods in the field are all based on supervised learning. This is a problem because supervised learning is done on synthetic datasets, resulting in poor generalization to real data. We propose a variational model for single-image dehazing based on a smooth approximation of the dark channel prior. We also formulate an unsupervised deep learning model for dehazing based on the same functional. We also propose a variational model for single-image rain removal. The proposed models do not require synthetic data, so they have better properties on real hazy and rainy images than the state-of-the-art models. We compared the proposed methods to the state-of-the-art on real images degraded by weather, and our methods proved to be better in restoring the images. The comparison was made subjectively by visual inspection and objectively by comparing the improvement in object detection performance of restored images. Contributions: Optimization based method for single color image haze removal based on variational formulation of the assumption on minimal value of color component and a biased object depth estimator Optimization based method for single image rain removal with loss function based on absolute value of directed gradient and sparsity constraints on the image of rain.

Abstract (croatian)

Broj uslikanih slika svakim se danom povećava. Mnoge od tih slika slikane su vani, gdje vrijeme može imati negativan utjecaj na kvalitetu slike. Degradacije uzrokovane vremenskim neprilikama mogu biti problematične, i za ljudskog promatrača i za efikasnost algoritama računalnog vida. Zato je problem obnavljanje takvih slika bitan. Ova disertacija se fokusira na obnavljanje slika degradiranih maglom ili kišom. trenutno najbolje metode se sve baziraju na nadziranom učenju. To je problem jer se nadzirano učenje temelji na sintetičkim skupovima podataka, što rezultira slabom generalizacijom na stvarne podatke. Razvili smo varijacijski model za uklanjanje magle baziran na glatkoj aproksimaciji pretpostavke o najtamnijem kanalu. Također smo formulirali metodu baziranu na nenadziranom učenju koristeći isti funkcional. Razvili smo i varijacijsku metodu za uklanjanje kiše sa slike. Razvijeni modeli ne trebaju sintetičke podatke za treniranje, pa imaju bolj rezultate na stvarnim maglovitim i kišnim slikama od do sada najuspješnijih modela. Usporedili smo predložene modele sa trenutno najuspješnijim modelima u području subjektivno, vizualnom inspekcijom rezulatat i uspoređivanjem efikasnosti tih modela kao alat za predprocesiranje u problemu detekcije objekata. Naši modeli su u svim slučajevima na stvarnim podatcima bili uspješniji. Doprinosi: Optimizacijska metoda za uklanjanje magle iz slike u boji temeljena na varijacijskom obliku pretpostavke o minimalnoj vrijednosti komponenti boje i pristranom procjenitelju udaljenosti objekta od kamere. Optimizacijska metoda za uklanjanje kiše iz slike u boji s funkcijom cijene koja je temeljena na apsolutnoj vrijednosti usmjerenog gradijenta slike i uvjetu rijetkosti slike koja sadrži samo kišu.

• Keywords: image restoration; variational methods; image dehazing; image deraining
• Keywords (croatian): obnavljanje slika; varijacijske metode; uklanjanje magle; uklanjanje kiše
• Language: english
• URN:NBN: urn:nbn:hr:168:405344
• Promotion: 2023
• Study programme: Title: Computer Science; Study programme type: university; Study level: postgraduate; Academic / professional title: Doktor znanosti (Doktor znanosti)
• Type of resource: Text
• Extent: viii, 88 str. : ilustr. ; 31 cm.
• File origin: Born digital
• Access conditions: Open access
• Created on: 2023-03-31 09:45:15