HiG > IMT > master > thesis > proposals MT

Thesis proposals MT

P1: Bridging the gap between features and semantic meanings

Background

Content-based image retrieval (CBIR) has been a very active research field in the last decade. A multitude of research papers were published and a number of retrieval systems were proposed both from the academy and industry. The shear majority of the proposed techniques, however, operate on low-level features of the image content. Therefore they fail to capture the high-level conceptual meanings conveyed by the scene; represented as a combination of regions, colors or textures in an image. Furthermore, the seldom approaches combining two or more low-level features do it mainly to avoid complex analysis of the image content such as image segmentation. Hence, they ignore the semantic meanings of such combination. The human users of CBIR systems are used to use high-level concepts to describe their targets rather than low-level features such as colors, textures, edges, frequencies, etc. This in turn led to a growing interest among the CBIR research community in retrieving images based on high-level concepts and semantic meanings. To achieve this goal, there is a need to bridge the semantic gap, i.e., to find algorithms that will infer high-level semantic concepts (sites, objects, and events) from low-level image features that can be easily extracted from the content (color, texture, shape and structure, layout, motion, and audio pitch and energy).

Objectives

The objective of this thesis project is the investigation of the approaches extracting features both from the color contents and its spatial distribution to further relate the features to the scene contents semantic meanings. The implicit goal is to make the computers mimic the human way of analyzing the image content and use it to retrieve images from large image databases. The work will consist of:

Faculty Contact

Faouzi Alaya Cheick, faouzic(at)hig.no

P2: Highlights extraction from unscripted video

Background

Extracting highlights from sports video is more challenging than extracting them from scripted content e.g. from a movie. This is due to the unreliability of shot detection in unscripted content. For example, in a football game video, the visual features are very similar in most of the video frames; which makes grouping them into different shots a difficult task even when there are clear semantic units within the same period such as attacks on the goal, and counter attacks in the mid-field. Additionally to the video track sports videos contain audio track which has been ignored in the early years of video retrieval by content. The audio track contains generally valuable information on the action in the video and therefore can be of a great help in achieving a better temporal segmentation of the video content. The association of complementary cues from the audio to those extracted from the video may reduce the number of false alarms in the detected markers and enables detection of highlight candidates that are of a higher semantic level. For instance, in a football game when the ball goes into the net (visual marker) is associated with a referee whistling (audio marker) that follows it, the corresponding segment is highly likely to be remarkable or interesting because very likely there was a goal scored in that segment. Hence, the detection of audio and visual markers enables a representation that is at a higher semantic level than that achieved only with visual markers.

Objectives

The objective of this thesis project is the development of a framework for the analysis of unscripted video content, e.g. sports videos, lecture videos, security videos, etc., which allows the detection of specific audio and visual markers and their use to extract the highlights from the videos. These highlights will be used to brows the unscripted video content.

The work will consist of:

Faculty Contact

Faouzi Alaya Cheick, faouzic(at)hig.no

P3: Unscripted video skimming

Background

Unscripted digital video is rapidly becoming important and can be encountered in several aspects of our daily life e.g. education, entertainment, security, law-enforcement, etc. The size of these video collections is also rapidly growing larger and becoming more difficult to manage and even to consult. Therefore, there is an urgent need to effectively browse segments of it in a short time without missing the global idea conveyed by content of the video. One approach to achieve this goal is to create a summary of the video by extracting the most significant audio and video information and create a "skim" video which represents a very short synopsis of the original. The association of complementary cues from the audio and video may reduce the number of false alarms in the detected markers and enables efficient extraction of highlight candidates that are at a higher semantic level. These highlights represent important segments of the video, thus they may be used for building a summary or condensed representation; while still keeping the essence of the original video.

This project is closely related to the work on "P2: Highlights extraction from unscripted video" and can be conducted in a single larger project.

Objectives

The objective of this thesis project is the development of a framework for unscripted video skimming. These highlights extracted from the unscripted video will be used with the audio-visual cues to build a hierarchical summary of the video material.

The work will consist of:

Faculty Contact

Faouzi Alaya Cheick, faouzic(at)hig.no

P4: Using eye-tracking for evaluating visually complex TV programming

Background

Traditionally, TV production results in a single stream of images - possibly with some static graphical overlay. Increasingly, TV programming becomes visually more complex. In many cases, TV programming is presented in ways that resemble web user interfaces with several areas/frames containing dynamic content. In SMS-TV, for instance, there is one frame for the main video stream, one for chat, several for text streams, etc; in sports, there may be several frames displaying parallel information streams; in political debates, there may be frames showing viewer comments and responses. There is currently a lack of knowledge of what these more visually advanced formats means to the user experience and the effectiveness of the producer-viewer communication.

Objectives

Eye-tracking equipment has been a useful tool for researchers for understanding where viewers direct their attention when interpreting user interfaces, text documets, or images. Eye-tracking equipment may also be a useful tool for deriving new insight into how users experience visually complex TV programming. This project will be a part of the TiDE (Television in a Digital Environment) project at HiL and shall result in new insight with regards to the following questions:

Faculty Contact

Rune Hjelsvold, rune.hjelsvold(at)hig.no and Frode Volden, frode.volden(at)hig.no

P5: Adaptive Learning Management Systems

Background

The Internet and the World-wide web are becoming channels for facilitating "anytime, anywhere" access to learning content. Learning Management System (LMSs) are tools for teachers and students alike for managing such learning content. Most LMSs today offer a "one size fits them all"-strategy for delivering learning content to the students. In order to be useful to individual learners, LMSs must be adaptive, i.e. the presentation of the course material must be adapted to the student.

Objectives

Apropos Internet, a Hamar-based company that developes solutions for the Internet, is in the process of transforming its e-learning platform, SmartLearn, to becoming an adaptive learning management system. The objective of this project is to study, test and compare solutions for:

Faculty Contact

Rune Hjelsvold, rune.hjelsvold(at)hig.no

P6: Lecture Videos On-demand

Background

Previous master's theses at HiG has investigated automatic analysis, segmentation and table-of-content generation for lecture videos. These projects have suggested algorithms based on analyzing the contents of the blackboard. These algorithms need to be refined and improved to become building blocks in Lecture video on-demand systems.

Objectives

The objectives of this project is to:

Faculty Contact

Rune Hjelsvold, rune.hjelsvold(at)hig.no

P7: Finding and Retrieving Media Objects in Tagged Multi-user Collections

Background

Traditionally, three different approaches have been taken for retrieving media objects from large collections: Controlled vocabularies allow the user to pick from predefined terms/keywords (possibly selected from rigidly defined ontologies) associated with the objects; fulltext retrieval systems allow the user to type in words or phrases that should appear in a textual description of the objects; content-based retrieval systems allow the user to describe the media objects in terms of content-based features (such as color, texture and shape for images, for instance). Semantic tagging is an alternative approach that is becoming popular — especially in systems that promote sharing of media objects owned by individual users (see Flickr, the photo sharing system owned by Yahoo, for instance). Web-based tagging systems allow participants to annotate a particular media object with a freely chosen set of keywords (tags). Tagging systems make it easy for the tagging user to choose the most best-fit terms for her own objects but makes it hard for novice and querying users to find the best-fit terms to use when formulating a query.

Objectives

The objective of this project is to investigate problems that novice and querying users experience and to develop techniques for improving the user experience for novice users and for users querying across collections stored on the system. The project may follow one or more of these directions:

Faculty Contact

Rune Hjelsvold, rune.hjelsvold(at)hig.no

P8: Utilizing Accurate Geospatial data in Interactive 3D Presentations

Background

Hafjell Alpine Centre and Hunderfossen Family park participate in a project called Digital Destinations. Their goal is to become leading experience destinations in Norway in utilizing media technology. One of the project activities is to develop interactive 3D presentations for customers to enjoy before and after a visit to the destination. Detailed geospation data has been collected from the destinations and can, for instance, be used to generate accurate 3D models of the downhill slopes at Hafjell. The data may also be used to generate interactive 3D presentations. The amount of data, however, is not suitable for such use.

Objectives

The objective of this project is to study, test and compare solutions for:

Faculty Contact

Rune Hjelsvold, rune.hjelsvold(at)hig.no

P9: Colour gamut mapping - how the GBD affects the GMA

Background

The set of all colours reproducible by a device is denoted its colour gamut. In colour imaging, one is faced with the problem of non-matching colour gamuts. This often makes it necessary to perform mapping of colours from a source to a reproduction gamut. An algorithm performing this task is referred to as a gamut mapping algorithm (GMA). An important part of a GMA is the definition of the boundary of the gamut, the so-called gamut boundary descriptor (GBD).

Objectives

Previous gamut mapping research at the Norwegian Color Research Laboratory has resulted in the development of a very powerful tool for gamut visualisation and gamut mapping, namely ICC3D (http://www.colorlab.no/icc3d/). It has also been demonstrated in a previous master project that the GBD can have a significant impact on the performance of a GMA. The proposed project will further investigate this relationship. This will be done through

If this experiment is successful, and particularly if the choice of GBD turns out to have a significant effect, the results of this project should be publishable in international conferences and scientific journals.

Faculty Contact

Ivar Farup (ivar.farup at hig.no), Arne Magnus Bakke (arnembak at online.no), and Jon Y. Hardeberg (jon.hardeberg at hig.no).

P10: Perceptual image difference thresholds in display and/or print

Background

The CIE colour difference formulae have been widely used in the image reproduction industries. However, these current formulae are designed to calculate the colour difference between two solid colour patches under a fixed viewing condition. In real image reproduction workflows, the images do not consist of large solid patches, and the visual appearance of images is influenced by various parameters, such as illumination, state of adaptation, visual sensation, and media type.

Objectives

There have been previous experiments to define the perceptible and/or acceptable threshold of perceived image difference, but there are still a number of undefined areas, such as the influence of independent factors and combination of factors (colour, noise, background etc.). In the proposed master project, the student will investigate the influence of selected factors in selected media (such as LCD and CRT displays, and print) through the use of a carefully designed psychophysical experiment.

Faculty Contact

Eriko Bando (eriko.bando at hig.no) and Jon Y. Hardeberg (jon.hardeberg at hig.no).

P11: Colorimetric and spectral characterization of digital cameras

Background

In order to use a digital camera to capture images of high colorimetric precision and quality, typically a colorimetric characterization procedure needs to be carried out. One example of this is the creation of an ICC profile using off-the shelf colour management tools together with a colour target containing a relatively high number of coloured patches. Typically a relationship between the camera's RGB colour space and a device-independent colour space such as CIELAB is found using regression techniques. However, in practice a camera is typically used under widely varying lighting conditions, and on scenes with different spectral characteristics than those of the colour target. For these and other reasons, the colorimetric precision is often compromised.

An alternative way to characterize a camera requires knowledge about its spectral sensitivity functions to establish a reliable mathematical model of the camera. If these sensitivity functions are known, more reliable estimates of the scene colour can be made.

Objectives

In the proposed project advantages and disadvantages of different target-based and spectral methods for characterizing cameras should be investigated qualitatively and quantitatively. An important tool in this research will be a newly acquired system to measure the spectral sensitivity using a device known as a monochromator.

Faculty Contact

Jon Yngve Hardeberg, jon.hardeberg(at)hig.no

P12: Multispectral image acquisition using an extended color filter array: a feasibility study

Background

Typical digital color image acquisition systems use CCD sensors with an RGB color filter array, that is, each pixel is sensitive to either red, green, or blue light. A full RGB image in which each pixel contains RGB information is calculated using different techniques known as demosaicing. This process can lead to artefacts such as color moiré.

Multispectral image acquisition usually requires multiple subsequent shots, typically though different filters in a filter wheel, or through a Liquid Crystal Tunable Filter. This, however, imposes the serious limitation that the scene needs to be stationary.

Objectives

To allow for fast acquisition of moving scenes, it could be possible to construct a CCD-sensor with an extended color filter array, containing not only the conventional red, green, and blue filters, but an extended set of spectrally selective filters.

In this project it would be convenient to use Matlab including the ImagEval toolbox (http://www.imageval.com/).

Faculty Contact

Jon Yngve Hardeberg, jon.hardeberg(at)hig.no

P13: Analysing the quality of color halftoning algorithms

Objectives

You will in this project study the halftoning of a color image for a CMYK printer. Using appropriate quality metrics you will compare the halftoning by error diffusion and halftoning by vector error diffusion. The goal is to evaluate how different or close are the two methods. To complete the study, the visual comparison of the prints by both method will done.

Faculty Contact

Jérémie Gerhardt (jeremie.gerhardt at hig.no) and Jon Y. Hardeberg (jon.hardeberg at hig.no)

P14: Digital restoration of movies

Background

Restoration of historical cinematographical archive material is a very important task. Unfortunately, film based media become unstable over time, unless they are stored at low temperatures and the humidity is controlled. Usually, a bleached colour release print is the only available record of a movie. Since the bleaching phenomenon is irreversible, photochemical restoration of faded prints is not possible, hence the incontestability of digital colour restoration.

Today's tools and technologies for digital restoration are to a large extent based on manual work, hence it is a very time consuming, costly, and variable process which depends a lot on the skills of the operator. Two identified problems in this process are that there are no references to how the movie actually appeared when it was new, and that there does not exist a framework for consistent communication between experts regarding the choices that are made when restoring a movie.

Objectives

The goal of the project is to develop and enhance tools for digital movie restoration, for example by introducing color management and semi-automatic spatial image processing methods such as the Retinex-based ACE algorithm. The project could be carried out in collaboration with the Norwegian Film Institute and international an international research partner.

Faculty Contact

Jon Yngve Hardeberg, jon.hardeberg(at)hig.no

P15: Electronic programme guides (EPG)

Background

Digital TV broadcasts often contains information on coming programs, but the different channels and channel providers have implemented this differently. Other sources of programme information are on the net and the different channels Teletext stream. When redistributing channels eg. over IP or cable, a nice additional service could be to provide an electronic programme guide that combines these sources of information. This project will be done in cooperation with the Norwegian company SnapTV that works with distribution of TV signals over IP.

Faculty Contact

Frode Volden, frode.volden(at)hig.no

P16: Personal channels – Low cost TV channels

Background

IP distribution of TV signals allows for low cost distribution of personal video content to other viewers. Personal Video-blogs or local television channels can relatively easy be made available to other audiences, but there are unsolved issues involved. Micro payment systems and security management of the content are among these unsolved issues. This project will be done in cooperation with the Norwegian company SnapTV that work with distribution of TV signals over IP.

Faculty Contact

Frode Volden, frode.volden(at)hig.no

P17: Disabilities and computer gaming

Background

Many people with disabilities wants to play computer games, but have difficulties finding games they can use due to the nature of their disability. A Norwegian company, MediaLT, have worked on a project where they have looked at requirements etc. for the user interface of computer games, to make them playable for people with various types of disabilities. Adaptations of games needs implementation and evaluation, and there is also room for new ideas on how to improve the gaming experience for people with various disabilities.

Faculty Contact

Frode Volden, frode.volden(at)hig.no

Additional Topics

10.05.2007
© Høgskolen i Gjøvik,
Postboks 191, Teknologivn. 22, 2802 Gjøvik, Tlf. 61135100, Faks 61135170, E-post: postmottak@hig.no
Ved spørsmål eller kommentarer til websiden, webredaksjonen

English