BSc / MSc Project Suggestions

 

Dr Pengwei Hao

Office : Engineering Building E104, School of EECS

Tel: 0044 20 7882 5207

email: p.hao@qmul.ac.uk

 

Currently, I have the following suggestions:

1)      Animated video generation for dental health

2)      Redeveloping a dental health website

3)      Computer games for dental health

4)      Video database management system

5)      A management system for video retrieval

6)      Video summarization/skimming for video surveillance

7)      Ambiguous Image Generation

8)      Harmony Index of Images

9)      Dimensionality reduction for hyperspectral data

10)  3D shape matching and error estimation (co-supervised with Dr Lifong Zou, the QM Institute of Dentistry)

11)  Perfect 3D surface data (co-supervised with Dr Lifong Zou)

12)  Progressive 3D mesh coding

 

 

1) Animated video generation for dental health

 

Development of a series of animated videos to promote dental health; including healthy diet, tooth-brushing, what causes tooth decay etc. Some dental pictures can be used for the videos. The animation can be 2-dimensional using 2D graphics or 3-dimensional using 3D graphics, and some cartoon or animation or graphics tools and multiple techniques can be used for the video generation.

The same cartoon characters could be involved in all the videos to provide continuity.

 

 

 

2) Re-developing a dental health website

 

Re-developing the website design to make it more user-friendly and easier to navigate - alter appearance of menus. Multiple techniques can be used for the development, such as HTML, flash animation. The webpages should be opened nicely on computers or mobile phones or tablets.

See www.toothloop.com and http://www.nhs.uk/Conditions/Dental-decay/Pages/Introduction.aspx

 

 

 

3) Computer games for dental health

 

Development of a computer game with several levels to try and educate children about good dental health. Characters used in the educational videos could also appear in the game. Two ideas are:

a.       A game where you have to look after a character and their teeth and keep them healthy - involves choosing the right foods, taking them for dental check ups, brushing their teeth

b.      A game where a character travels through an environment e.g. city, jungle etc, and has to pick up healthy foods, avoid bacteria, remember to stop to brush their teeth etc, and would be a points based game. The character could progress to different levels. The concept is similar to that of Sonic the hedgehog.

The games can be run on computers or mobile phones or tablets.

 

 

 

4) Video database management system

 

To design a user-friendly GUI for video database management with menu, large/medium/small iconic display of video clips, video information display, add/update/delete/split/merge video clips or information. Video information includes when and where the video was taken, length of the video, some annotations that users or providers added. Video clips are provided as individual video files. Video display can be used to view single video clip in a subwindow, to parallelly view multiple video clips with alignment in multiple subwindows, to successively and seamlessly view multiple video clips in a subwindow, and to view from a given time point and at different speeds, e.g. 1/8x, 1/4x,1/2x, 1x, 2x, 4x, 8x,16x the normal speed.

 

The outcomes should include:

1)      A user-friendly GUI

2)      A programme for video file and information management and manipulation

3)      An integrated programme with video database management for video display for various requirements

 

 

 

5) A management system for video retrieval

 

To design a user-friendly GUI management system for video retrieval, with menu, large/medium/small iconic display of video clips, video information display, add/update/delete/indexing/search video clips or information. Video information includes when and where the video was taken, length of the video, some annotations that users or providers added. Video clips are provided as individual video files. By video indexing/retrieval, video clips should be sorted and displayed in multiple subwindows, and could be viewed parallelly with alignment in multiple subwindows and and viewed from a given time point and at different speeds, e.g. 1/8x, 1/4x,1/2x, 1x, 2x, 4x, 8x,16x the normal speed.

 

The outcomes should include:

1)      A user-friendly GUI

2)      A programme for video and video information indexing/retrieval

3)      An integrated programme with video indexing/retrieval for various requirements

 

 

 

6) Video summarization/skimming for video surveillance

 

Video summarization/skimming is to identify high-value information events in a video stream and to present it to a user. This project is for surveillance video summarization/skimming using detection and clustering of important events. Assuming that events are main source of energy change between consecutive frames set of interesting frames is extracted and then clustered. Based on the structure of clusters, video summaries can be created static and dynamic. Static summary is build of key frames that are organized in clusters. Dynamic summary is created from short video segments representing each cluster and is used to lead user to the event of interest captures in key frames. Some experiments are needed for the implementation.

Reference:

U. Damnjanovic, V. Fernandez, E. Izquierdo, J.M. Martinez, "Event Detection and Clustering for Surveillance Video Summarization", Ninth International Workshop on Image Analysis for Multimedia Interactive Services(WIAMIS), 7-9 May 2008, pp. 63 - 66.

 

The outcomes should include:

1)      A user-friendly GUI

2)      A programme for video skimming, especially for simple surveillance video clips

3)      A programme for video skimming with improved performance and comparable to other programmes available on web

 

 

 

7) Ambiguous Image Generation

 

Ambiguous images are visual illusion images, crafted to exploit graphical similarities and other properties of visual system interpretation between two or more distinct image forms due to the phenomenon of multistable perception. Ambiguous images were previously created purposefully by artists, such as Salvador Dali in his paintings. With recent development in visual psychology and image processing, it is now possible for us to develop algorithms to generate ambiguous images. We can start with some simple constraints and simple ambiguity. For example, the outline of the ambiguity can be given, then we can find some objects in a big segmented dataset whose outline is similar to the given, and finally we can reshape the retrieved image objects and fill into the defined outline.

 

 

Summary:

Problem: Find the matched curves and images

Methods: Curve registration, image processing and image retrieval (similarity measuring)

Resource: The Berkeley Segmentation Dataset (http://www.eecs.berkeley.edu/Research/Projects/CS/vision/bsds/), etc.

 

 

 

8) Harmony Index of Images

 

Digital cameras are so popular and cheap that almost everyone can become a photographer. However, not all of us are professional, and the photos are often badly taken. How can we tell a camera what a good picture is? We need find some features to compute and give a number for image quality the assessment - harmony index - how harmonious a good picture is. It should depend on nice-looking colour/intensity distribution, beautiful object geometry, etc. However, some bad photography may be difficult to find, such as a headless people picture.

 

 

Summary:

Problem: Find theoretically and numerically what the photographic problem is.

Methods: Measuring the harmony index of colour/intensity distribution, object geometry, but headless people may not be easy.

 

 

 

9) Dimensionality reduction for hyperspectral data

 

Hyperspectral images may have tens or hundreds of spectral bands, i.e. a big number of pixel values at each spatial point in an image, see the images below. In such images, the data between bands are highly correlated. To process or use such images easier and faster, we wish to reduce the number of the spectral bands to a small number of representative spectral bands or several new representative spectral bands with mixed original bands. Dimensionality reduction methods can be used achieve this, such as linear transforms by principle component analysis (PCA).

 

Hyperspectral image of 80 bands with roads,water and rice paddy

The interspectral data are highly correlated

 

Summary:

Problem: Can we find some representative spectral bands? Or several “mixed bands”?

Methods: Component analysis and transforms

 

 

 

10) 3D shape matching and error estimation

Co-supervisor: Dr Lifong Zou, the Institute of Dentistry, Queen Mary, University of London, email: l.zou@qmul.ac.uk

 

3-D co-ordinate data can be digitized through numerous data acquisition devices. With advanced technologies in electronics and mechanics, an integration of CAD/CAM in dentistry is no longer a dream. This project of the fitness analysis of a crown and a crown preparation is a tool of assessment for a new CAD/CAM application in dentistry. We wish to know how much difference between two data sets. To find the difference, we have to first register the two data sets so that the points of the two data sets can be aligned and then the error between them can easier found and visualized.

 

The data formats from these devices are vary. It is often happens, a data format conversion has to be done before the data analysis. Therefore, this project is also involved with data format conversion.

 

crown-coping3

The inner surface of the crown

The outer surface of the crown preparation

The fitness analysis of the inner surface of a crown and the outer surface of a prepared tooth

 

Summary:

Problem: The fitness analysis of the inner surface of a crown and the outer surface of a prepared tooth

Methods: 3D point set / mesh matching

 

 

 

11) Perfect 3D surface data

Co-supervisor: Dr Lifong Zou, the Institute of Dentistry, Queen Mary, University of London, email: l.zou@qmul.ac.uk

 

3-D surface data can be acquired with 3D scanning devices, see the images below, but the surfaces may be too complex to scan in a single run. There always is some error or gap between the data of two scans. To perfect the acquired 3D data, we have to remove the redundant imperfect points and find a better-shaped point set. In other words, we should stitch of 3D point sets / meshes so that the new data set looks perfect.

 

Acquired imperfect 3D data

An example of better shaped 3D data

 

Summary:

Problem: Remove the redundant imperfect points and find a better-shaped point set

Methods: Stitching of 3D point sets / meshes

 

 

 

12) Progressive 3D mesh coding

 

3D mesh coding methods are important for efficient storage and fast transmission of large models. Such polygon models come from a number of sources including computer-aided design, range scanners, terrain mapping and iso-surface extraction from volume data. As model sizes continue to grow, finding efficient and improved compression methods will remain an important problem in computer graphics. In particular if complex scenes are built in 3D and sent over communications lines and need to be animated in real time the compression methods become crucial for the quality of animation to be achieved. A particular application where this is true is the computer games played interactively over the net and shared immersive worlds. Virtual worlds require transmission of massive amounts of triangulated 3D geometric data over the network. For both efficient transmission and real-time rendering of models to and within the end-user computers at various levels of detail, there was a need for an efficient technique of mesh simplification and progressive compression. There have been three aspects of current research: mesh simplification, single-resolution mesh compression, and progressive mesh reconstruction, see pictures below.

Mathematically, we have proved that the upper bound of the bit rate of 3D mesh coding, 3 bits per vertex. However, the experiments have to be done to support that. Through experiments, we also wish to know how much is the gap between the theoretical result and the practical results. The method we suggest to use is face contraction/refinement, see picture below. We have some previous version of the source codes for edge contraction in C, but the codes must be changed for the new idea. (Our previous experiments with edge contraction give 4.18 bits/vertex in average, while the theoretical upper bound is 4.85 bits/vertex. The previous technical report can be downloaded here.).

 

Single-rate / Progressive transmission of a compressed 3D mesh

 

Face contraction/refinement for 3D mesh coding

 

Summary:

Problem: The lower bound of bit rate of triangular meshes has been proved mathematically, but how close is it to the practical rates?

Methods: Use face removal to encode a mesh progressively and give the bit rates for lossless decoding.