Improving quality of scanned visual content using convolutional neural networks

Toroman, Jovan (author), Čehovin Zajc, Luka (mentor)

In this work we approach the problem of image quality improvement for images of documents captured using smartphones. Our goal is to make content captured this way as similar as possible to the original groundtruth images. We make a twofold contribution, (1) we propose an innovative method for improving quality of documents using convolutional neural networks and (2) we create a training dataset containing images captured using smartphones under dierent external conditions (lighting, viewing angle). This dataset is captured under controlled external conditions using an acquisition setup developed for this purpose. In our work we use six dierent smartphones and one hundred groundtruth images. We build our work on two different existing convolutional neural network architectures, UNet and DPED network, using them as our starting point. Both models are adapted to our domain. We experiment with dierent hyperparameters for both networks, as well as with dierent forms of training. We evaluate results of the two variants of our method against each other and also against other baseline approaches (simple contrast enhancer and a function from Adobe Photoshop Express). We use standard image quality comparison metrics to objectively compare performance. From the results we can see that our work outperforms baseline approaches, especially in more dicult scenarios of uneven illumination. Finally, we discuss the results of our method and possible improvements.

• Language: English
• Keywords: captured document quality enhancement, smartphone scanner, convolutional neural networks
• Work type: Master's thesis/paper
• Typology: 2.09 - Master's Thesis
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2020
• PID: 20.500.12556/RUL-122425
• COBISS.SI-ID: 42265859
• Publication date in RUL: 10.12.2020

Secondary language

• Language: Slovenian
• Title: Izboljšava kvalitete zajetega slikovnega gradiva s konvolucijskimi nevronskimi mrežami

Abstract

V tem delu se lotevamo problema izboljšanja kakovosti slik slikovnih dokumentov, posnetih s pametnimi telefoni. Naš cilj je dobiti slike čim bolj podobne prvotnim referenčnim slikam. Predlagamo dva prispevka prispevka, (1) predlagamo inovativno metodo za izboljšanje kakovosti slik s konvolucijskimi nevronskimi mrežami, in (2) ustvarimo podatkovno zbirko, ki vsebuje slike, posnete s pametnimi telefoni v različnih zunanjih pogojih (osvetlitev, kot gledanja). Slike so zajete v nadzorovanih zunanjih pogojih z uporabo naprave, razvite za namen tega dela. Za zajem smo uporabili šest različnih pametnih telefonov, zajeli smo sto različnih izvornih slik. Naša delo temelji na dveh konvolucijskih arhitekturah nevronskih mrež, UNet in DPED. Oba modela sta prilagojena za uporabo v naši problemski domeni, preizkusimo več kombinacij hiperparametrov ter načinov učenja. Rezultate obeh različic naše metode primerjamo med seboj in tudi glede na dva referenčna pristopa (preprost ojačevalec kontrasta in rešitev, dostopna v programu Adobe Photoshop Express). Za primerjavo uporabljamo standarne mere za objektivno ocenjevanje podobnosti slik. Iz rezultatov vidimo, da naša metoda deluje bolje kot referenčne metode, še posebej v težjih pogojih z neenakomerno osvetlitvijo. V zadnjem delu rezultate tudi komentiramo in izpostavimo možnosti za nadaljnje delo.

• Keywords: izboljšava kakovosti slik, optični čitalnik za pametne telefone, konvolucijske nevronske mreže