Search

Ovaj se rad bavio problemom kreiranja sustava za detekciju promjene televizijskog sadržaja, tj. klasifikacijom slika televizijskog sadržaja u dva razreda: sadržaj i promjena sadržaja. Pokazalo se da rješenje ovog problema nije trivijalno te zahtijeva korištenje dubokog učenja. Vizualni transformer opisan u radu „An Image is Worth 16x16 Words“ pokazao se korisnim u rješavanju navedenog problema. U radu je prvo prezentiran detaljan opis problema nakon kojeg slijedi opis korištenog skupa podataka. Nakon toga je dan teoretski pregled dubokog učenja i učenja nad sekvencijskim podacima. Zatim je detaljno objašnjena implementacija sustava, a na kraju se nalazi poglavlje s postignutim rezultatima. This paper addresssed the problem of creating a system for detecting changes in television program, i.e., classifying television program into two categories: program and program change. It was found that solving this problem is not trivial and requires the use of deep learning. The vision transformer described in the paper „An Image is Worth 16x16 Words“ proved to be useful in solving the aforementioned problem. The paper first presents a detailed description of the problem, followed by a description of the dataset. After that, a theoretical overview of deep learning and sequence models is provided. The implementation of the system is then explained in detail and the paper concludes with a chapter presenting the achieved results.

Procjena osvjetljenja bitan je faktor u mnogim primjenama računalnog vida. Cilj je pronaći vektor osvjetljenja kako bi se odradila transformacija slike čiji bi rezultat bila slika s najviše čovjeku vidljivih informacija. U ovom radu provjerava se metoda zasnovana na ideji klasifikacije slika ovisno o tipu svjetlosti koji je dominantan u slici, kao koraku koji prethodi samoj procjeni osvjetljenja. Iznesena je teoretska podloga potrebna za izgradnju takvog modela i opisani napravljeni koraci. Konačni rezultat je izlučen korištenjem Cube+ skupa podataka te VGG i EfficientNet arhitektura kao temelja za dobivanje modela sa zadovoljavajućim rezultatima. Color constancy is an important factor in many computer vision applications. Its primary task is to find an illumination vector such that when an image is transformed with it, it appears as tough it was illuminated with a white light source, thus increasing the amount of information the image transfers to humans. This work uses a method based on prior image classification depending on type of scene illumination source, as a step of preprocessing when estimating illumination of images. Theory behind the process of building such a model was explained and steps described. The final result is extracted using Cube+ dataset coupled with VGG and EfficientNet architectures as a basis for reaching a model with satisfying evalutation results.

Računalna postojanost boja problem je odstranjivanja utjecaja osvjetljenja scene u slici. Većina današnjih metoda bavi se procjenom globalnoga osvjetljenja, a problem procjene više osvjetljenja u sceni do sad je neistražen. U ovom radu razvijena je metoda za segmentaciju slika scena s više osvjetljenja temeljena na dubokom učenju. Glavni cilj segmentacije je prepoznati dijelove slike koji su osvijetljeni direktnim odnosno ambijentalnim izvorom osvjetljenja. Zatim je moguće iskoristiti tu informaciju za korekciju slike s obzirom na lokalna osvjetljenja. Opisane su metode segmentacije i procjene globalnoga osvjetljenja potrebne za rješenje problema. Predloženi model za segmentaciju sastoji se od poznatih arhitektura poput U-Net modela i VGGNet modela. Computational color constancy is a problem of eliminating the effect of a scene illuminant in an image. Most of today’s methods deal with estimating global illumination, and the problem of estimating multiple illuminants in a scene has so far been unexplored. In this thesis, a method for image segmentation of multi-illuminant scenes based on deep learning has been developed. The main goal of segmentation is to identify parts of the image that are illuminated by a direct or ambient light source. Then it is possible to use this information to correct the image with respect to local illumination source. Segmentation and estimation of global illuminant methods needed to solve the problem are described. The suggested segmentation model consists of known architectures such as the U-Net and the VGGNet.

Duboke neuronske mreže su vrlo moćni alati za vizualnu analitiku, odnosno interpretaciju i analizu slika, te su pokazale vrhunske performanse u raznim zadacima, pa tako i u onome obrađenom ovdje, a to je detekcija voznih traka. U usporedbi s tradicionalnim modelima plitke računske strukture, jedna vrlo važna prednost neuronskih mreža je ta što su reprezentacije podataka izgrađene u proces učenja automatski. Shodno tomu, neuronske mreže se smatraju sposobne za naučiti razlikovati sve značajke od interesa, naglašavajući da je ručna konstrukcija obilježja zamijenjena automatskim učenjem. Najčešće korištene mreže za detekciju cestovnih traka su konvolucijske neuronske mreže (CNN). U metodi za detekciju traka predstavljenoj ovdje se, koristeći duboko učenje, istrenirao model konvolucijske neuronske mreže tako da pronalazak lijeve i desne linije na cestovnih trakama bude robusniji i brži od modela koji se temelji na tradicionalnim algoritmima, odnosno na računalnom vidu. Dobiveni rezultati ukazuju na moć modela konvolucijskih neuronskih mreža koji je sposoban s velikom točnošću i malom funkcijom gubitka vrlo dobro odrediti trake i riješiti zadani problem. Deep neural networks are a really powerful tool for visual analytics, for image interpretation and analysis, and have shown some really good performances in various tasks, including the one discussed here, which is lane detection. In comparison to traditional models with shallow computational structure, a very important adventage of neural networks is that the representations of the datasets are built in the process of learning automatically. Accordingly, neural networks are considered to be capable of learning to make a difference between all of the features of interest, emphasizing that the manual construction of these features is replaced by the autonomous learning. Most commonly used neural networks for lane detection are convolutional neural networks (CNN). In lane detection method proposed here, the CNN model was trained using deep learning, so that detecting of left and right line of the road lane is more robust and faster than from the model based on traditional algorithms. The results point out the power of CNN models that are capable of detecting lanes and solving problems with a big accuracy value and small loss function value.

Detekcija objekata jedan je od aktualnih i složenih problema računalnog vida. Detektori objekata mogu se podijeliti na one koji koriste predefinirane sidrene okvire i one koji ih ne koriste. Ovaj rad detaljno opisuje rad neuronskih mreža i konvolucijskih neuronskih mreža koje se koriste u analizi slike. Nakon toga opisuje se arhitektura nekih poznatijih detektora koji ne koriste predefinirane sidrene okvire. To su FCOS, CornerNet i CenterNet. FCOS je uzet kao predstavnik ove skupine, a Faster R-CNN kao predstavnik detektora koji koriste predefinirane sidrene okvire. Performanse ova dva detektora testirana su i uspoređena na DOTA skupu podataka. Object detection is currently one of the most complex computer vision tasks. Object detectors can be divided into two groups: anchor-based detectors and anchor-free detectors. This thesis describes in detail neural networks and convolutional neural networks used in image analysis. After that, it also describes the architecture of some of the most well-known anchor-free detectors: FCOS, CornerNet and Center- Net. FCOS was chosen as representative of this group and Faster R-CNN was chosen as representative of anchor-based detectors. Performances of these two detectors were tested and compared on DOTA dataset.

Non-destructive ultrasound evaluation is a technique for flaw detection in materials. Among other things, it is used for monitoring key components of nuclear power plants, railways, and pipelines. Analysis of ultrasound data is performed by human inspectors manually analyzing acquired images. Such a process can be tedious and is highly dependent on the inspector's previous experience. Deep learning methods are hard to develop because of the lack of available data. The shortage also impacts training new human experts in this field, since they learn through experience. Data from real inspections can not be used because of non-disclosure agreements. On the other hand, blocks with synthetic flaws are expensive to produce. The goal of this work is to develop methods for defect detection. To improve the method’s performance, deep learning methods for generating additional synthetic images need to be developed. Generated data should be realistic and of high quality even to human experts in ultrasound evaluation. Additional data should be used to improve the performance of the deep learning defect detector. Although some methods for defect detection and classification were developed, they are mostly based on traditional image and signal processing techniques such as magnitude threshold. Such an approach does not yield satisfactory results and can not generalize well on data from various ultrasound probes and scanned materials. In the field of generating synthetic ultrasound data, yet much has to be done. Only some traditional and Finite Element Methods (FEM) have been proposed. All this means that there is much space for improvements and new ideas in the field of defect detection in ultrasound non-destructive evaluation. Guided by this need for improvements, in this thesis, several methods for non-destructive ultrasound data analysis and ultrasound image generation are proposed. These methods can be summarized as follows: * Generative adversarial network with additional object detector discriminator for synthesis of high quality realistic ultrasound images * Improved one-stage defect detector trained using additional ultrasound images synthesised using generative adversarial network * Method for anomaly detection in ultrasound images. Experimental results of the proposed methods are presented and discussed. The proposed generative method outperforms existing methods when the quality of images is assessed by human experts. The generated images can even be used to improve the performance of the state-of-the-art deep learning defect detector. Also, the analysis of different methods for anomaly detection in ultrasonic images is given alongside the improvements made to some methods.

Ultrasonic testing is a non-destructive evaluation (NDE) technique that is used to inspect the integrity of the material and check if there are any defects in its internal structure. The acquisition of ultrasonic data is already done in an automated fashion using robotic manipulators, but the analysis of the data is still done manually by specially trained experts. Manual analysis is subject to human errors especially when a large amount of data needs to be inspected. The goal of this doctoral dissertation is to develop deep learning-based methods that can be used to efficiently and reliably detect defects from ultrasonic images. Deep learning methods have been achieving great results in many image analysis tasks during the past few years. However, to develop a good deep learning-based method for defect detection, several problems must be solved. Some of the encountered challenges include a small dataset, noise and geometry signals, and extreme aspect ratios of the defects. Ultrazvučno testiranje je tehnika nerazornog ispitivanja koja se koristi za ispitivanje kvalitete materijala i detekciju defekata unutar njegove strukture. Akvizicija ultrazvučnih podataka se već obavlja automatizirano korištenjem robotskih manipulatora, ali analizu prikupljenih podataka još uvijek ručno rade obučeni inspektori. Takva analiza podložna je ljudskoj pogrešci, pogotovo ako je količina podataka koju je potrebno pregledati velika. Cilj ove doktorske disertacije je razviti metode bazirane na dubokom učenju koje bi se mogle koristiti za efikasnu i pouzdanu detekciju defekata na ultrazvučnim slikama. Metode dubokog učenja zadnjih godina postižu odlične rezultate u raznim problemima vezanim za analizu slike. Međutim, kako bi se razvila dobra metoda za detekciju defekata bazirana na dubokom učenju, nekoliko problema mora biti riješeno. Neki od izazova koji se pojavljuju su mali skup podataka, šumoviti podaci, pojava signala uzrokovanih geometrijom i ekstremni omjeri oblika defekata.

Optical coherence tomography (OCT) images of the retina provide a structural representation and give an insight into the pathological changes present in age-related macular degeneration (AMD). Due to the three-dimensionality and complexity of the images, manual analysis of pathological features is difficult, time-consuming, and prone to subjectivity. Computer analysis of 3D OCT images is necessary to enable automated quantitative measuring of the features, objectively and repeatedly. As supervised and semi-supervised learning-based automatic segmentation depends on the training data and quality of annotations, a new database of Annotated Retinal OCT Images (AROI database) has been created. It consists of 1136 images with annotations for pathological changes (fluid accumulation and related findings) and basic structures (layers) in patients with AMD. Inter- and intra-observer errors have been calculated to enable the validation of developed algorithms in relation to human variability. The framework for validation of OCT image segmentation methods consisting of a publicly available database of manually labeled image features and methodology and programming code for standardized measurement of segmentation accuracy is developed thus making the first scientific contribution. In the case of retinal OCT images of patients suffering from AMD, there is a significant deformation of the retinal structure. This causes the methods for retinal layers segmentation without introducing information on pathology significantly impaired. Therefore, an approach for layer segmentation which takes into account information about fluids (joint segmentation of layers and fluids) was chosen. Also, the automatic segmentation with standard U-net architecture and two state-of-the-art architectures for medical image segmentation has been performed to set a baseline for further algorithm development and to get insight into challenges for automatic segmentation. After an analysis of how various architectures and their modifications affect the performance and accuracy of segmentation, an analysis of the effects of various data manipulations (data augmentation, image post-processing) on segmentation performance was performed. Based on all the examined methods, comparison and proposal of two models for segmentation of retinal and intraretinal fluid layers in optical coherence tomography images that give comparable results were made. One model consists of a standard U-net architecture with post-processing using conditional random fields. The second model consists of a neural network architecture that combines the good features of the U-net architecture and the ResNet architecture. The existing state-of-the-art architecture has been adapted to the given problem and improved by adding skip connections. Additionally, segmentation accuracy was improved by using data augmentation and post-processing using conditional random fields. This is the second scientific contribution. A comparative quantitative analysis of fluid-only segmentation and fluid segmentation with available layer information was performed to gain insight into how much information about layers (fluid position in the retina) contributes to the accuracy of fluid segmentation itself. A study of the efficiency of transfer learning, which is an active and insufficiently researched topic in the field of deep learning with applications in medicine, was conducted, based on which a method for retinal fluid segmentation in optical coherence tomography images was proposed. This is the third scientific contribution. In conclusion, the main challenges of automatic segmentation are presented and the ideas for further research and improvements are given. Optička koherentna tomografija (OCT) slika retina omogućava strukturalni prikaz i daje uvid u patološke promjene prisutne kod senilne makularne degeneracije (SMD). S obzirom na trodimenzionalnost i kompleksnost slika, manualna analiza patoloških značajki je teška, vremenski zahtjevna i podložna subjektivnosti. Stoga je računalna analiza 3D OCT slika nužan preduvjet za automatsko kvantitativno mjerenje značajki, na objektivan način u sklopu kliničke prakse. Kako nadzirana i polu-nadzirana automatska segmentacija bazirana na dubokom učenju ovisi o podacima za treniranje i o kvaliteti anotacija, prikupljena je nova baza anotiranih retinalnih OCT slika (AROI baza). Sastoji se od 1136 slika pacijenata koji boluju od SMD s pripadajućim anotacijama patoloških promjena (akumulacije tekućina i slične patološke promjene) i osnovnih anatomskih struktura (slojeva). Izračunata je procjena ljudske varijabilnosti kako bi se omogućila usporedba i validacija razvijenih algoritama u odnosu na ljudsku varijabilnost. Time je razvijen radni okvir za provjeru metoda za segmentaciju OCT slika koji se sastoji od javno dostupne baze slika s ručno označenim značajkama te metodologije i od programskog koda za standardizirano mjerenje točnosti segmentacije čime je ostvaren prvi znanstveni doprinos. U slučaju retinalnih OCT slika pacijenata koji boluju od SMD dolazi do značajne deformacije strukture retine. Time je onemogućena segmentacija slojeva retine bez uvođenja informacija o patološkim promjenama bez značajnog narušavanja točnosti segmentacije razvijenih modela. Zato je odabran pristup segmentacije slojeva pri čemu se uzima u obzir informacija o tekućinama, odnosno istovremena segmentacija slojeva i tekućina. Provedena je automatska segmentacija sa standardnom U-net arhitekturom i dvije napredne i češće korištene arhitekture za segmentaciju medicinskih slika kako bi se postavila osnova za daljnji razvoj algoritama i dobio uvid u izazove automatske segmentacije. Nakon što je provedena analiza kako razne arhitekture i njihove modifikacije djeluju na točnost i performanse segmentacije, provedena je analiza djelovanja raznih manipulacija nad podacima (povećanje podataka, obrada slika) na performanse segmentacije. Na osnovu svih ispitanih metoda, napravljena je usporedba i prijedlog dvaju modela za segmentaciju slojeva retine i intraretinalnih tekućina u slikama optičke koherentne tomografije koji daju sumjerljive rezultate. Jedan model se sastoji od standardne U-net arhitekture uz obradu predikcija pomoću uvjetnih slučajnih polja (eng. conditional random fields). Drugi model se sastoji od arhitekture neuronske mreže koja kombinira dobre značajke U-net arhitekture i ResNet arhitekture. Postojeća napredna (eng. state-of-the-art) arhitektura prilagođena je za dani problem i unaprijeđena dodavanjem veza (eng. skip connections). Dodatno, točnost segmentacije je poboljšana uporabom augmentacije podataka (eng. data augmentation) i obradom predikcija pomoću uvjetnih slučajnih polja. Time je ostvaren drugi znanstveni doprinos. Napravljena je komparativna kvantitativna analiza segmentacije samo tekućina i segmentacije tekućina uz dostupne informacije o slojevima kako bi se dobio uvid koliko informacije o slojevima, odnosno poziciji tekućina u retini doprinose točnosti segmentacije samih tekućina. Provedeno je istraživanje efikasnosti prijenosnog učenja koje je aktivna i nedovoljno istražena tema u području dubokog učenja s primjenama u medicini na temelju čega je predložena metoda za segmentaciju intraretinalnih tekućina u slikama optičke koherentne tomografije. Time je ostvaren treći znanstveni doprinos. Zaključno su izneseni i istaknuti glavni izazovi automatske segmentacije te dane ideje za daljnja istraživanja i poboljšanja.

Ovaj rad bavi se računalnim stvaranjem slika oblaka dizelskog goriva iz raspršivača u motoru s unutarnjim izgaranjem, sa svrhom unaprijeđenja i istraživanja biodizelskog goriva. Time se nastoji riješiti problem ograničene količine dostupnih slika dobivenih snimanjem, kako je iz relativno malenog broja izvornih slika moguće digitalnim stvaranjem dobiti skup slika dostatne veličine za daljnje istraživanje. Uvode se i opisuju općeniti koncepti i metode dubokog učenja poput neuronskih mreža, te posebice GAN (Generative Adversarial Networks), skupina neuronskih mreža korištenih za stvaranje novih slika na temelju postojećih, pomoću koje je implementirana metoda stvaranja slika oblaka goriva. This paper explores the process of digitally generating images of fuel clouds dispersed in an internal combustion engine, with the purpose of advancing the research of biodiesel fuels. It thereby attempts to tackle the problem of the limited amount of actual recorded images, as image generation using real references can greatly enhance the number of images suitable for research. The general methods of deep learning are described, as well as neural networks including GANs (Generative Adversarial Networks), a family of neural networks used to generate high-quality images using reference, which were used to implement the method of image generation.

Nedostatak slika s defektima kod ispitivanja materijala bez razaranja predstavlja problem za automatizaciju detekcije defekata. Rad proučava uporabu dubokih generativnih mreža kao rješenje ovog problema. Cilj je ostvariti jeftiniji i pouzdaniji proces generiranja slika s defektima. The low amount of images with defects with non-destructive testing makes automating defect detection hard. This thesis studies the use of generative networks as a solution of this problem with the goal of developing cheaper and more reliable method of generating images with defect.

U sustavima računalnih vida --- bilo da se temelje na obradi videa ili statičkih slika --- česta je pojava degradacija uzrokovana prirodnim elementima. Jedna od takvih degradacija je uzrokovana kišom, koja znatno utječe na moderne sustave zadužene za praćenje objekta u pokretu. Počevši od te problematike, razvija se problem uklanjanja takve degradacije sa jedne slike, koja se ne može oslanjati na vremensku domenu u oklanjanju kišnih anomalija. Ovaj rad detaljno prolazi kroz koncepte i raznovrsne metode uklanjanja kiše sa slike temeljenih na dubokom učenju. Također prolazi kroz implementaciju i rezultate dvije od njih. In modern computer vision systems --- be they based on processing videos or static images --- a common occurrence is degradation of input data by natural elements. One such degradation is caused by rain, which severely impacts systems, such as the ones for motion detection. Using that as a starting point, a sub-domain of problem-solving is developed for removing rain from purely static images, which cannot rely on the existence of a time domain for removing anomalies caused by rain. This work goes into the key concepts and various methods of single-image rain removal, primarily based on deep learning. In addition, it details the implementation and results of two of such methods.

Razvoj računarskih znanosti uvijek je bio popraćen razvojem tehnologije. Sve češća upotreba modernih vozila zahtjeva visoke tehnološke standarde. Upravo je računalni vid grana koja uvelike pronalazi primjenu u tehnološkom napretku modernih vozila. Važan zadatak koji rješava računalni vid je procjena dubine scene. Korištenje podataka o dubini vozilima omogućuje izbjegavanje sudara s objektima za koje se utvrdi da su preblizu. Na taj način se može drastično utjecati na smanjenje prometnih nesreća i samim time spasiti mnoge živote. Cilj ovog rada bio je napraviti vlastito rješenje na temelju postojeće metode za procjenu dubine scene korištenjem samo jedne slike. U skladu s postavljenim ciljem, u radu su predstavljene umjetne neuronske mreže s posebnim naglaskom na duboke konvolucijske neuronske mreže. Opisan je korišteni kitti set podataka, kao i detalji implementacije same metode te su predstavljeni dobiveni rezultati. The development of computer science has always been followed up by the development of technology. The increasing use of modern vehicles requires high technological standards. Computer vision is part of artificial intelligence which has a wide use in technological advancement of the modern vehicles. The important task solved by computer vision is depth estimation of the scene. Usage of depth data allows vehicles to avoid collisions with objects that are believed to be too close, and that is how number of traffic collisions may be drastically reduced and many lives saved. The aim of this paper was to create my own solution based on existing monodepth method. In addition to the subject, this paper presents artificial neural networks with special accent on deep convolutional neural networks. Used kitti data set is described, as well as the details of implementation of the method itself, and the obtained results are presented.

Motorna vozila koja za svoje kretanje koriste motore s unutarnjim izgaranjem u cilindre motora ubrizgavaju gorivo u obliku spreja. Kako bi se ubrizgivanje spreja moglo optimizirati, uslikane su slike spreja goriva koji ulazi u cilindre. Slike su segementirane metodama dubokog ucenja i analizirane. Analiza uklju ˇ cuje odre ˇ divanje duljine penetracije spreja i kut ¯ raspršenja. Kako bi se slike segmentiranle, korištena je konvolucijska neuronska mreža koja se sastoji od konvolucijskih slojeva. Mreža je po arhitekturi U-Net i sastoji se od dva dijela cija je uloga opisana u radu. Motor vehicles that use internal combustion engines inject a spray of fuel into the cylinders of their engine. In order to optimize the spray injection, images of the fuel spray entering the cylinders were taken. Images are segmented using the deep learning methods and analyzed. The analysis includes determining the spray penetration length and spray angle. In order for images to be segmented, a convolutional neural network consisting of convolutional layers is used. The network is based on the U-Net architecture and consists of two parts whose role is described in the paper.

Problem koji ovaj rad opisuje i pokušava riješiti zasniva se na klasifikacijama slika s nepravilnostima i bez, uz uvjet da se na ispitni materijal/predmet ne nanosi šteta(NDT metode). NDT metode su obližnje opisane u uvodu zajedno s primjerima rješavanja sličnih problema korištenjem starijih metoda strojnog učenja. Kako se pristup rješavanja problema u ovom radu zasniva na dubokom učenju, objašnjene su osnove neuronskih mreža, dubokog učenja te konvolucijskih neuronskih mreža. U nastavku su objašnjene osnove dva široko korištena modela konvolucijskih neuronskih mreža primijenjenih u ovom radu, kao i jedan proizvoljno napravljeni model za usporedbu performansi. Objašnjena je programska izvedba na Google colabu kao i popratne tehnologije a kako je ispitni skup slika veoma bitan dio rada njima je također posvećeno poglavlje za detaljniji opis. Prije prikazivanja eksperimentalnih rezultata ukratko su opisane metrike uspješnosti te razlog korištenja odabranih. Rezultati su prikazani u obliku grafova i tablice nadopunjeni popratnim komentarima i bitnim dijelovima o procesu učenja modela. Na kraju je sve objedinjeno u zadnjem poglavlju zaključka i diskusije gdje razmatramo uspješnost rada, njegova ograničenja kao i moguće daljne korake. The problem that this paper describes and tries to solve is based on the classification of pictures with and without irregularities, provided that no damage is caused to the test material / object (NDT methods). NDT methods are described in more detail in the introduction along with examples of solving similar problems using older machine learning methods. Since the approach of problem solving in this paper based on deep learning, the basics of neural networks, deep learning and convolutional neural networks are explained. The basics of two widely used models of convolutional neural networks applied in this paper are explained below, as well as one arbitrarily constructed model for performance comparison. The program implementation on the Google Colab as well as the accompanying technologies are explained, and as the test set of images is a very important part of the work, a chapter for a detailed description is also dedicated to them. Before presenting the experimental results, metric performance was briefly described as well as the reason for using the selected ones. The results are presented in the form of graphs and tables supplemented by accompanying comments and essential parts of model learning process. Finally, everything is summarized in the last chapter, conclusion and discussion, where the success of the work, its limitations as well as possible further steps are discussed.

U ovom je radu opisana state-of-the-art metoda za segmentaciju slojeva retine i motivacija iza nje. Kroz prilagodbu metode domaćoj bazi podataka analizirane su performanse modela primijenjenog na bazu podataka pacijenata koji boluju od senilne makularne degeneracije. U sklopu rada provedeno je nekoliko eksperimenata te analiziran značaj pojedinih komponenti funkcije pogreške i težina pridijeljenih slojevima i granicama između slojeva retine. In this paper, the motivation behind the problem of retinal layer segmentation is explained and a state-of-the-art method for this problem is described. Through adjustment of the method to our dataset, the performance of the model was analysed when applied to a dataset of B-scans of patients suffering from age-related macular degeneration. A number of experiments were conducted as part of this paper and the importance of different loss function components and weights given to retinal layers and layer borders was analysed.

Segmentacija anatomskih struktura očne pozadine olakšava dijagnostiku i liječenje očnih bolesti. U diplomskog radu opisana je potpuno konvolucijska neuronska mreža za semantičku segmentaciju slojeva retine koja je temeljena na U-Net i EfficientNet arhitekturi. Eksperimenti provedeni u radu pokazali su kako odabrana arhitektura uspijeva postići bolje rezultate u usporedbi s najsuvremenijim metodama iz područja. Također, arhitektura je pokazala robusnost prilikom treniranja na slikama iz dva različita OCT uređaja. Dobri rezultati robusnih modela sugeriraju kako mreža uspijeva naučiti značajke koje razlikuju slojeve retine, te se ne oslanja na razlike u tehnikama snimanja između različitih OCT uređaja. Segmentation of the anatomical structures of the ocular background facilitates the diagnosis and treatment of ocular diseases. This master's thesis has described a fully convolutional neural network for semantic segmentation of retinal layers based on U-Net and EfficientNet architecture. Experiments conducted in this thesis have shown that the selected architecture achieves better results compared with the state-of-the-art methods in the field. Also, the architecture has shown robustness when training on images from two different OCT devices. Good results from robust models suggest that the network manages to learn features that distinguish retinal layers and does not rely on differences in imaging techniques between different OCT devices.

Posljednjih godina, sve se više istražuje područje računalno potpomognute medicinske dijagnostike. Za veliko poboljšanje u rezultatima problema detekcije i segmentacije, zaslužan je strelovit razvoj metoda dubokog učenja. U ovom radu fokus je na problemu segmentacije tekućina na OCT slikama mrežnice koristeći nove vrste umjetnih neuronskih mreža, kapsularne neuronske mreže. Dan je pregled dosadašnjih ostvarenja na području segmentacije OCT slika mrežnice i kapsularnih neuronskih mreža. Za potrebe rada, u Keras-u je implementirana relativno nova varijacija kapsularnih mreža, Matwo-CapsNet. Performanse implementacije testirane su na podskupu RETOUCH baze podataka OCT slika mrežnice. Dobiveni su vrlo dobri rezultati, mjerljivi s ponajboljim metodama koje su sudjelovale u originalnom RETOUCH izazovu. In the last few years, there has been an increasing interest in the field of computer-assisted medical diagnostics. The rapid development of deep learning methods is responsible for the great improvement in the results of detection and segmentation problems. In this paper, the focus is on the segmentation of fluids on retinal OCT images using new types of artificial neural networks, capsule neural networks. Firstly, we present an overview of previous achievements in the fields of segmentation of OCT retinal images and capsule neural networks. Secondly, we present the implementation of a relatively new variant of capsule networks, Matwo-CapsNet, that has been implemented in Keras for the purposes of this paper. Testing was performed on the subset of RETOUCH database of OCT retinal images. Acquired results are very satisfactory, comparable to the results of best methods participating in the original RETOUCH challenge.

Razvoj tehnologije sa sobom direktno povlači razvoj mnogih drugih područja kao što je medicina. Nedavnim ubrzanim razvojem strojnog i dubokog učenja, računala uspijevaju izvršavati visoko apstraktne zadatke iznad razine ljudskih sposobnosti. Ovaj rad opisuje segmentaciju medicinskih OCT slika mrežnice oka koristeći duboko učenje te uspoređuje dvodimenzionalni i trodimenzionalni pristup tom problemu. Development of modern technology causes development of other fields of study such as medical sciences. With the recent fast-paced development of machine and deep learning, computers are able to perform highly abstract tasks on levels higher than human capability. This paper proposes a solution for segmentation of medical OCT images of the retina and also compares two-dimensional approach with the three- dimensional approach to that problem.

Detekcija osoba u video zapisima je važna u mnogim realnim primjenama uklju- čujući autonomna vozila. U okviru diplomskog rada potrebno je istražiti metode za detekciju pješaka u prometnim scenama snimljenim iz vozila. Naročito treba proučiti metode koje su temeljene na strojnom učenju. U okviru diplomskog rada treba odabrati metodu za detekciju pješaka korištenjem duboke neuronske mreže i koju je potrebno programski implementirati. Učenje mreže treba provesti korištenjem skupa ispitnih video zapisa. Evaluaciju točnosti metode detekcije pješaka treba provesti korištenjem video zapisa koji nisu bili korišteni za učenje. Pedestrian detection is extremely important in many areas, especially autonomous vehicles. Focus of this paper is to explore and study current methods for detecting pedestrian in scenes captured from vehicle camera, with focus on machine learning algorithms. A method of choice will be implemented using deep neural networks. Chosen neural network has to be trained on video frames and evaluated on a separate set of video frames that network has not been trained on.

U nuklearnoj industriji vrlo je bitno u što kraćem roku naći eventualne defekte u materijalu, kako bi nuklearna elektrana što prije nastavila s radom. Takod̄er, izuzetno je bitno koristiti metode nedestruktivnog ispitivanja, s obzirom na to da su dijelovi jako skupi, a jedna od najpopularnijih je ultrazvučno ispitivanje. Te defekte traže ljudi, no med̄utim vrlo se rijetko neki nad̄u, a sam proces je vrlo mukotrpan. Stoga se koriste duboke neuronske mreže, kako bi se negativni primjeri što prije eliminirali. C-snimci su vrlo korisni za analizu, s obzirom na to da otkrivaju „3. dimenziju“ defekata, te se stoga ovaj Rad bavi automatskom detekcijom defekata korištenjem RetinaNet modela s ResNet50 backbone-om. Dobiveni rezultati imaju prostora za napredak, vjerojatno zbog prevelike kompleksnosti RetinaNet modela, te je potrebno daljnje istraživanje u ovom području. In the nuclear industry it’s very important to find eventual defects in the material as soon as possible, to keep the halting of the nuclear power plant to the minimum. In addition to that, nondestructive evaluation methods are critical, due to the price of the parts used, and one of the most popular ones is ultrasonic testing. These defects are human-searched, but they are very rarely found, and the process is very exhausting. That’s why deep neural networks are being used, to eliminate the negative examples quickly. C-scans are very useful for analysis, due to the fact that they reveal "the 3. dimension" of defects, and that’s why this Thesis is doing automatic detection using the RetinaNet model with ResNet50 backbone. The end results have space for improvements, probably because the RetinaNet model is too complex, and therefore further research is necessary.

U ovome radu opisan je postupak klasifikacije srčanog tkiva uporabom dubokog učenja. Opisani su podatci s kojima se radilo te način obrade podataka. Kako bi se povećao skup podataka, korišteni su postupci augmentacije podataka te su i oni detaljno opisani u radu. Za klasifikaciju srčanog tkiva koristila se konvolucijska neuronska mreža prilagođena iz VGG16 mreže te je proveden prijenos učenja. Opisan je postupak treniranja i prikazani su rezultati mreže kroz niz metrika. Dan je primjer idejnog sustava za klasifikaciju za kojeg je također provedeno ispitivanje. Grad-CAM algoritam je korišten za vizualizaciju rada mreže i primjeri vizualizacija su prikazani u radu. Na posljetku je objašnjen postupak segmentacije slika tkiva koristeći Grad-CAM vizualizaciju kao temelj algoritma. This paper describes a method of classifying cardiac tissue using deep learning. It describes the data that was used and how the data was processed. It also describes data augmentations that were used to increase the amount of available data. Convolutional neural network based on VGG16 architecture is used to classify cardiac tissue and transfer learning is applied. The training process is described as well as the results of the testing process using numerous metrics. It gives an example of a classification system that was also tested. GradCAM algorithm is used to visualize the workings of the neural network and the examples of visualizations are displayed. Finally, it describes a tissue segmentation process using GradCAM visualizations as the basis of the method.

The purpose of this study is to give a performance comparison between several classic hand-crafted and deep key-point detector and descriptor methods. In particular, we consider the following classical algorithms: SIFT, SURF, ORB, FAST, BRISK, MSER, HARRIS, KAZE, AKAZE, AGAST, GFTT, FREAK, BRIEF and RootSIFT, where a subset of all combinations is paired into detector-descriptor pipelines. Additionally, we analyze the performance of two recent and perspective deep detector-descriptor models, LF-Net and SuperPoint. Our benchmark relies on the HPSequences dataset that provides real and diverse images under various geometric and illumination changes. We analyze the performance on three evaluation tasks: keypoint verification, image matching and keypoint retrieval. The results show that certain classic and deep approaches are still comparable, with some classic detector-descriptor combinations overperforming pretrained deep models. In terms of the execution times of tested implementations, SuperPoint model is the fastest, followed by ORB. The source code is published on \url{https://github.com/kristijanbartol/keypoint-algorithms-benchmark}.

In the case of structured data, such as 2D images, many variants of traditional convolution neural network architectures have been successfully proposed. Learning from unstructured sets of data, such as sets of 3D point clouds, is a challenging task due to numerous reasons among which two most important ones are: 3D point cloud is generally (i) unordered and (ii) sparse data set. Therefore, the architectures have been proposed which are invariant to both ordering and number of points in the point cloud. PointNet is one such architecture, originally introduced and demonstrated on the task of classification and segmentation of the ModelNet40 data set. In this work we study the performance of PointNet on an even more demanding task, segmentation of human body parts. Finding enough training data of enough quality is generally a problem in deep learning, and especially for human body segmentation. To that end we take advantage of SMPL model which provides human body models in many shapes and sizes in an essentially automatic fashion, therefore avoiding a cumbersome procedure of manual collection and preparation of training data. Our results show that the proposed PointNet variant trained using SMPL model provides competitive segmentation results on the task of human body segmentation.

Posljednjih godina su metode dubokog učenja u mnogim slučajevima ostvarile su- periorne rezultate za probleme segmentacije slika u odnosu na tradicionalne metode. Modeli ispitani u ovom radu, CNN-S i U-Net su duboke konvolucijske neuronske mreže čija je arhitektura dizajnirana za segmentaciju slojeva retine na slikama dobive- nima optičkom koherentnom tomografijom. CNN-S koristi regresijski model, a U-Net klasifikacijski. Oblik i debljina slojeva retine su ključni za dijagnozu brojnih bolesti oka. Slike nad kojima su trenirane obje mreže prikazuju retine pacijenata koji boluju od senilne makularne degeneracije. Ovaj rad detaljno opisuje pripremu podataka za treniranje. Zatim je izložena korištena implementacija obje mreže da bi se konačno usporedili i objasnili validacijski rezultati pomoću standardnih statističkih mjera toč- nosti. In recent years deep learning based methods have demonstrated superior results in image segmentation problems over traditional methods. The models investigated in this study, CNN-S and U-Net are deep convolutional neural networks whose arhitec- ture is designed for segmentation of retinal layers in images acquired through optical coherence tomography. CNN-S implements regression model, while U-Net imple- ments classification model. Shape and thickness of retinal layers are very important features for diagnosing various eye diseases. Images used for training both networks show retinal layers taken from patients that suffer from age-related macular degenera- tion. This thesis describes data preprocessing in detail. Afterwards, implementations of both networks is presented. Finally, the validation results of two networks are shown and compared using standard statistical measures.

Današnji digitalni fotoaparati imitiraju ljudsku sposobnost ispravljanja boja u slici pomoću implementiranih algoritama za računalnu postojanost boja u svom sustavu za obradu slike. Postoje razne metode kojima se to pokušava ostvariti, a u ovom radu je predložena metoda učenja pod nadzorom koja korištenjem dubokih konvolucijskih neuronskih mreža nad dva skupa ulaznih slika (Cube i Cube+) i njihovog kanonskog osvjetljenja (engl. ground truth) ispravlja boje u slici. Rezultati eksperimenta su pokazali visoku točnost nad skupom ispitnih slika većinom slikanih po danu (Cube), dok su se rezultati nad zahtjevnijm skupom slika, koji uključuje i slike noćnih scena (Cube+), pokazali nešto lošijima u odnosu na dosadašnje provedene metode. Today's digital cameras imitate human ability of color correction using implemented color constancy algorithms in their image processing system. There are a variety of methods to achieve this, and this paper suggests a supervised learning method that uses deep convolutional neural networks for image color correction on two image datasets (Cube and Cube+) and their ground truth illumination. The results of the experiment showed high precision on the Cube dataset (daily scenes), while the results obtained on a more complexed Cube+ dataset (which includes some night scenes) showed lower precision compared to the previously tested methods.

Jedna od glavnih značajki današnjeg doba su velike količine nestrukturiranih podataka koji su raspoloživi za korištenje, ali ne mogu se u potpunosti iskoristiti radi ograničavajućih faktora u procesima označavanja podataka. U isto vrijeme u kontekstu dubokog učenja, pokazalo se kako korištenje većih količina podataka za treniranje modela daje bolje rezultate u smislu preciznosti, uz pretpostavku da se koristi model koji ima zadovoljavajući kapacitet. Ne bi li premostili prepreku nedostatka označenih podataka tako da mogu iskoristi što više dostupnih podataka, znanstvenici se danas ponovo vraćaju klasičnim metodama računalnog vida i kombiniraju ih s principima dubokog učenja te na taj način omogućuju neuronskim mrežama da uče tek uz manju pomoć nadziranog faktora. Ovo se protivi principima potpuno nadziranog učenja gdje model treba eksplicitno naučiti funkciju koja preslikava parove ˇ (ulaz, izlaz). Ovaj rad je demonstracija kako je korištenjem anotacija za klasifikaciju slika moguce izvršiti i semanticku segmentaciju te, implicitno i lokalizaciju objekata na slici. U ovu svrhu, razvijena je nova metoda za vizualizaciju regija pozornosti u neuronskim mrežama treniranih s regijama pozornosti koja ne uzima u obzir samo regiju pozornosti posljednjeg sloja neuronske mreže, vec i regije pozornosti svih ostalih (unutrašnjih) slojeva. Prvi rezultati ukazuju kako je kombiniranjem regija pozornosti iz unutrašnjih slojeva neuronske mreže moguće dobiti regiju pozornosti veće rezolucije nego u prethodnim metodama. Rezultati su također usporedivi s trenutno vodećim metodama u slabo nadziranoj semantičkoj segmentaciji slika, čak i bez eliminacije šuma. Today's era is marked by large amounts of unstructured data that is readily available, but cannot be fully utilized due to the bottlenecks in the annotation processes. Coincidentally, so far in the context of deep learning, it has been shown that the more data is used for training, the higher accuracy can be achieved, assuming that a reasonably complex model is used. To mitigate the annotation gap and make use of as much of (unannotated or weakly-annotated) data for training as possible, scientists today are reverting back to classical computer vision approaches and embedding their findings into deep learning models which are then capable of making sense of data with some help, as opposed to the fully-supervised approaches where the model is constrained to learn a mapping function given the predefined (input, output) pairs. This thesis demonstrates how image-level labels that are used for image classification can serve as weak annotations for semantic segmentation and, implicitly, object localization. For this purpose, a novel attention visualization technique for guided attention inference networks that considers the network’s inner-layer attention rather than only the penultimate convolutional layer’s attention is developed. Preliminary proof of the concept results show that combined inner-layer attention maps have higher resolution (i.e. object boundaries appear sharper in the image) than in the previous approaches and are comparable to state-of-the-art in weakly-supervised semantic segmentation even without addressing the noise issue.

Dijabetička retinopatija je jedna od glavnih kroničnih bolesti i jedan od glavnih uzroka sljepoće koja se može spriječiti u svijetu. Kako bi se postigla rana dijagnoza puno truda se mora uložiti u sustave automatskog probira pacijenata temeljene na slikama očne pozadine. Ova doktorska disertacija se bavi istraživanjem naprednih metoda obrade i analize slika koje su potrebne za razvoj automatskih sustava probira pacijenata. Prvi znanstveni doprinos je baza od pedeset slika očne pozadine koja sadrži slike zdravih, ali i osoba koje imaju dijabetičku retinopatiju. Baza slika sadrži i oznake normalnih kao i patoloških struktura. Slike su označila pet oftalmologa te se baza koristi za razvoj i validaciju algoritama. Drugi znanstveni doprinos je metoda za detekciju krvnih žila u slikama očne pozadine temeljena na modeliranju i višerazinskom praćenju krvnih žila. Kako bismo locirali optički disk koji se nalazi u svakoj slici očne pozadine razvijena je metoda za lociranje optičkog diska temeljena na glasanju i stohastičkom učenju te ta metoda predstavlja treći važni doprinos ove disertacije. Unutar disertacije pokazujemo kako ovaj pristup daje bolje rezultate od individualnih algoritama koji su dio ansambla algoritama. Kako bismo detektirali eksudate koji su jedan od najvažnijih prvih simptoma prilikom dijagnoze dijabetičke retinopatije razvijena je metoda koja kombinira izlaz duboke konvolucijske neuronske mreže sa specifičnim oftalmološkim znanjem unutar ekspertnog sustava. Ta metoda predstavlja četvrti znanstveni doprinos disertacije. Na kraju disertacije dajemo pregled rada te se daju prijedlozi za poboljšanje performansi sustava. Diabetic retinopathy is one of the leading disabling chronic diseases, and one of the leading causes of preventable blindness in the world. In order to achieve early diagnosis of diabetic retinopathy a major effort will have to be invested into automatic screening systems using color fundus photographs. This thesis investigates advanced image processing and analysis methods, which are needed for automatic screening system development. The first contribution of this thesis work is a database of fifty fundus images from healthy and diabetic patients. The database has normal and pathological structures labeled by five ophthalmology experts and is used for algorithm development and testing. The second contribution is a method for blood vessel segmentation from fundus photographs using model-based multi-scale vessel tracking. In order to locate the optic disc, which is present in all fundus photographs, a method based on a votingbased classifier and stochastic learning is presented as one of the thesis contributions. We show that this approach easily outperforms methods which are part of the classifier ensemble. In order to detect exudates, which are one of the most important early signs of diabetic retinopathy, a method based on combining a deep convolutional neural network with specific ophthalmic knowledge in one expert system was developed and represents the last thesis contribution. In the end, the thesis gives a summary of the work with considerations about potential performance improvements.

Non-destructive ultrasonic testing (UT) of materials is used for monitoring critical parts in power plants, aeronautics, oil and gas industry, and space industry. Due to a vast amount of time needed for a human expert to perform inspection it is practical for a computer to take over that task. Some attempts have been made to produce algorithms for automatic UT scan inspection mainly using older, non-flexible analysis methods. In this paper, two deep learning based methods for flaw detection are presented, YOLO and SSD convolutional neural networks. The methods' performance was tested on a dataset that was acquired by scanning metal blocks containing different types of defects. YOLO achieved average precision (AP) of 89.7% while SSD achieved AP of 84.5 %.

Ovaj rad bavi se detekcijom umora vozača uporabom dubokih umjetnih neuronskih mreža. Opisana je motivacija za rad na sprječavanju posljedica od umora vozača i pregled metoda koje su trenutno korištene u postojećim vozilima. Objašnjeno je što je to duboko učenje i osnove konvolucijskih neuronskih mreža. Opisana je metoda koja koristi duboko učenje i treptanje kao indikator umora. Prikazana je i implementacija programa temeljenog na toj metodi i osnovni alati koji su bili nužni da se ona ostvari. Na kraju su prikazani rezultati primjene implementacije na skupu podataka i predložena su poboljšanja i naputci za daljnji rad. This paper discusses drowsiness detection in drivers using deep artificial neural networks. It presents the motivation for preventing drowsy driving and an overview of methods being currently used in existing vehicles. In this work, an explanation is given on what deep learning is and the basics of convolutional neural networks. It then describes a method which uses deep learning and blinking as a drowsiness indicator. It also shows a software implementation for the described method and tools which were necessary for that implementation. At the end of the paper results of implementation are shown on a dataset and suggestions are given for further improvements.

U ovom radu prikazana je ukratko povijest računalnog vida s fokusom na duboke konvolucijske neuronske mreže. Napravljena su i testirana dva sustava za detekciju i praćenje ljudi na slici i videu: prvi je Viola-Jones algoritam temeljen na kaskadama Haarovih značajki, te YOLO (You Only Look Once) algoritam, jedan od trenutno najboljih algoritama za detekciju objekata temeljen na konvolucijskim neuronskim mrežama. Na kraju su opisani dobiveni rezultati te su analizirani krivo prepoznati uzorci i situacije. In this work, brief history of computer vision is shown. The focus is on the deep convolutional neural networks. Two systems for people detection and tracking are built and tested: first one is Viola-Jones algorithm based on Haar features cascades, and YOLO (You Only Look Once) algorithm, one of the best state-of-the-art algorithms for object detection based on deep convolutional neural networks. Following that, numerical results are expressed from the classifiers, their performance is evaluated and misclassified examples and situations are briefly discussed.

Sustavi za automatsko čitanje tablice imaju razne primjene, od sustava sigurnosti i kontrole prometa do olakšavanja parkinga. Postupak se sastoji od dva koraka: detekcije tablice na slici te raspoznavanja teksta na tablici. Detekcija tablice na slici je složen problem analize slike. U okviru rada su istraženi modeli za detekciju objekata temeljeni na dubokom učenju. Implementirano je rješenje koristeći Faster R-CNN metodu. Sustav je učen na generiranom skupu slika tablica te je model evaluiran i analizirane su njegove mogućnosti. Reading license plates automatically can have many uses, from security systems and traffic control to easier parking regulation. The system consists of two parts: license plate detection and optical character recognition. License plate detection is a complex image analysis problem. Several deep learning detection models are analyzed. Implemented solution uses Faster R-CNN. The system is trained on a generated dataset, evaluated and analyzed.

Raspoznavanje teksta je važna primjena analize slike te se primjenjuje u mnogim područjima ljudske djelatnosti. Istraživanja su pokazala da pristupi temeljeni na dubokom učenju daju visoku točnost u mnogim područjima. U okviru ovog diplomskog rada proučene su primjene metoda dubokog učenja za višeobjektnu detekciju, a jedna od njih, \textit{Faster R-CNN} je iskorištena u problemu raspoznavanja teksta. Predložene su brojne prilagodbe i poboljšanja modela kojima pristup znatno nadmašuje klasične metode u raspoznavanju teksta. Text recognition is an important application of image analysis with applications in many real life situations. Studies have found that approaches based on deep learning achieve excellent results in many areas. Deep learning methods for multiple object detection have been examined within this master's thesis. One of them, \textit{Faster R-CNN} is used for text recognition problem. Many adaptations and improvements of the approach have been submitted. With these modifications the approach excels classsic approaches in text recognition problem.

U ovome radu ispitali smo mogućnosti treniranja Faster R-CNN duboke neuronske mreže nad generiranim podacima kroz slučaj detekcije i klasifikacije znakova na ispisanim računima. Fokus rada bio je na analizi stvarnih podataka te implementaciji sustava za njihovo automatsko generiranje. Analizom rezultata ustanovljena je visoka sličnost generiranog i stvarnog skupa podataka po pitanju mogućnosti lokalizacije znakova, dok je kod klasifikacije uočen znatniji data mismatch odnosno veća razlika generiranog i stvarnog skupa podataka. Predloženo objašnjenje tog problema jest manjak određenih fontova u generiranom skupu podataka, odnosno neki su fontovi prisutni samo u stvarnom skupu, no ne i u generiranom. In this thesis we explored the possibilities of training Faster R-CNN deep neural network on automatically generated dataset, investigating the case of detection and classification of characters on receipts. The focus of the thesis was to analyze real data and to implement a system for its automatic generation. In case of network’s localization capabilities, high similarity between generated and real dataset were detected, meaning low data mismatch was achieved. When it comes to classification, the data mismatch is significantly higher. A proposed explanation of that problem is the lack of certain fonts in the generated dataset, meaning there are some fonts present in the real dataset, but not in the generated one since they are not easily obtainable.

Ovaj rad daje usporedbu efikasnosti Faster R-CNN modela i klasičnih metoda računalnog vida za problem detektiranja igrača u scenama. Ispitani modeli Faster R-CNN-a učeni su na dva različita skupova s označenim pješacima. Prikazani su rezultati, navedeni problemi i nedostaci te mogućnosti dorade. Klasičan pristup zasniva se na oduzimanju pozadine te dodatnom procesiranju detektiranih regija. Korištene su paralelne i optimirane funkcije iz biblioteke otvorenog koda OpenCV. Ispitani su parametri k-NN metode oduzimanja pozadine i morfoloških transformacija za otklanjanje uvjeta scene. Opisani su algoritmi praćenja i analize detektiranih igrača te rezultati i moguća poboljšanja. In this thesis we compare efficiency of Faster R-CNN model and classical methods of computer vision for detecting players in soccer scenes. Tested models of Faster R-CNN were trained on two different pedestrian datasets. Results, potential problems and improvements were presented. Classical approach is based on background subtraction and further processing of detected regions. Parallel and optimized functions from OpenCV library were used. Parameters of k-NN background subtraction method and morphological transformations for scene conditions were tested. Algorithms for tracking and player analysis, and results and possible improvements are described.

U ovom radu provedeno je istraživanje primjene dubokih konvolucijskih neuronskih mreža na problemu raspoznavanja znakova. U tu svrhu razvijen je sustav za učenje i testiranje različitih arhitektura neuronskih mreža. Podsustav za učenje implementiran je pomoću razvojnog alata za paralelno računanje na grafičkoj kartici kako bi postupak bio što brži. Dio sustava za testiranje razvijen je s ciljem detaljne analize rada naučenih modela, te jednostavne i pregledne usporedbe više različitih modela. Istraživanje je ostvareno uz dodatnu podršku tvrtke Microblink specijalizirane za razvoj tehnologija računalnog vida za mobilne uređaje. Za učenje i testiranje modela koriste se od tvrtke interni skupovi podataka, a prilikom testiranja provodi se analiza i njihovog internog klasifikatora. Rezultati su prikazani i komentirani sa ciljem pronalaska optimalne arhitekture modela za primjenu na mobilnom uređaju. Pritom je osim točnosti naučenih klasifikatora uzeta u obzir i njihova kompleksnost te prosječno trajanje klasifikacije. In this thesis, deep convolutional neural networks for optical character recognition have been researched. For this purpose a system for learning and testing of different architectures of neural networks has been developed. Subsystem for learning was implemented with development tools for parallel computing on graphics cards to make the procedure as fast as possible. Part of the system for testing was developed for detailed analysis of the learned models, as well as simple and transparent comparison of different models. The research had additional support from the company Microblink specialized in the development of computer vision technology for mobile devices. Internal datasets from the company are used for the learning and testing of models, and additional analysis was conducted on the companies internal classifier. The results are presented and discussed with the aim of finding the optimal model architecture for the application on a mobile device. In addition to the accuracy of the learned classifiers, their complexity and average classification duration was taken into account as well.