From pet@bruce.csse.monash.edu.au Tue Dec 2 13:22:20 2003 Received: from bruce.csse.monash.edu.au (localhost [127.0.0.1]) by bruce.csse.monash.edu.au (8.12.8+Sun/8.12.8) with ESMTP id hB22MK8x005439 for ; Tue, 2 Dec 2003 13:22:20 +1100 (EST) Received: (from pet@localhost) by bruce.csse.monash.edu.au (8.12.8+Sun/8.12.2/Submit) id hB22MKCX005438 for annn@csse.monash.edu.au; Tue, 2 Dec 2003 13:22:20 +1100 (EST) Date: Tue, 2 Dec 2003 13:22:20 +1100 (EST) From: Peter Tischer Message-Id: <200312020222.hB22MKCX005438@bruce.csse.monash.edu.au> To: annn@csse.monash.edu.au Subject: honours projects from Peter TISCHER for 2004 X-Spam-Status: No, hits=0.0 required=5.0 tests= version=2.20 X-Spam-Level: Status: RO Content-Length: 9510 Lines: 188 Tuesday, December 2, 2004 Ann, Here are some ideas for honours projects for 2004. The proposals are in the format Bernd Meyer requested for this year. Peter TISCHER RETAIN 1: Generalised Vector Median Filtering 2: Peter TISCHER 3: Median filtering is a way of removing impulsive noise from images whose pixel values are scalar. The median operation is defined for scalars but it needs to be generalised when we want to talk about the median of a set of vectors. Colour images have pixels which have a red, green and blue scalar components. Other images whose pixels take on vector values include multispectral images and fMRIs. The project will involve coming up with a framework for generalising the concept of median filtering to images with vector valued pixels and studying the performance of these generalised median filters on a variety of image types but with special interest in performance on colour images and fMRI sequences. 4: BCS 24 pt, could suit BDS 24 pt as programming will not be difficult 5: References - none 6: Special Prerequisites - none, student need not have taken 3rd year Image Processing +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ REMOVE 1: Pixel Interpolation for Picture Enlargement 2: Peter TISCHER 3: When we want to enlarge a picture we need to come up with values for pixels which are between pixels for which we already have values. A case where we are specially interested in this is where we have a high definition television and we want to display standard television pictures at a higher resolution so the emphasis is on being able to interpolate pixel values quickly. Coming up with new pixel values based on the values of surrounding pixel values is something which is at the heart of lossless image compression. In lossless image compression, this is known as prediction. The aim of the project is to show how existing pixel interpolation techniques are special cases of prediction techniques from lossless image compression and to investigate the trade-offs between interpolation quality and computational effort required when a variety of pixel prediction techniques are used. 4: BSE 12 pt, could suit BDS 24 pt as programming will not be difficult 5: References - none 6: Special Prerequisites - none, student need not have taken 3rd year Image Processing +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ RETAIN 1: Approximating frequency distributions by mixture models 2: Peter TISCHER 3: Often when we are trying to estimate the probability associated with a set of events we compute the relative frequency distribution. A question we may ask is whether the relative frequency distribution can be adequately described as a mixture of probability distributions where the components in the mixture are in some sense simple probability distributions. If we can can find a good approximating mixture model we may be able to show that the events can be thought of as originating from different sub-populations of the overall population. Separating things into sub-populations is important in data mining and machine learning. This project aims to come up with good mixture models for a given relative frequency distribution by using MML techniques. In the first place, the problem will be approached by treating it as a lossless data compression problem. This can be seen as a simpler form of MML. Should time permit, the problem might be approached using MML in its general form. 4: BCS 24 pt, 5: References - none 6: Special Prerequisites - none, It would be an advantage to have a maths background and to be taking the honours unit on 'Learning and Prediction'. +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ NEW for 2004 1: Improving the Near-Lossless Compression mode of JPEG-LS 2: Peter TISCHER 3: JPEG-LS is an international standard for the lossless compression of grey-scale images. It achieves good compression at high throughput rates. JPEG-LS also has a near-lossless mode. This achieves greater compression but introduces some reconstruction error. By specifying a parameter, the maximum reconstruction error in a pixel value can be controlled. The near-lossless mode in JPEG-LS is relatively inefficient with respect to compression in situations where the resulting image is actually quite compressible. In addition, the prediction strategy used in JPEG-LS is not well suited to near-lossless compression. The aim of the project will be to take an implementation of JPEG-LS and improve its performance with respect to near-lossless image compression by both improving its coding strategies and its prediction strategies. In addition, the project will investigate incorporating region of interest coding. In some parts of an image the user may want to tolerate no reconstruction errors while in other parts the user may specify quite large bounds on the reconstruction error. While the near-lossless work will aim at a wide range of image types, the region of interest coding may choose to concentrate more on medical images. 4: BSE 12 pt, BDS 24pt, BCS 24pt 5: References - none 6: Special Prerequisites - none +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ NEW for 2004 1: Document Segmentation for Document Image Compression 2: Peter TISCHER 3: The JBIG-2 standard for compressing binary images, that is images whose pixel values are either black or white, allows for images to be segmented and for different segments to be treated differently in the compression process. The standard, however, does not specify how this segmentation should be carried out. It specifies only how the encoder is to tell the decoder what the segmentation is. The aim of this project is to investigate ways of segmenting binary images of documents, such as those that arise in fax and document processing systems, so that the resulting segmentation may be used to encode the binary image most efficiently. The segmentation will therefore be aimed at finding regions that have similar properties with respect to how compactly the information in those regions may be represented. Thus, an image of a page might be broken into segments which represent text, headlines, tables, half-tone images or white space at the margins of a page. The project might also investigate how to use such a segmentation to get best compression of the binary image. 4: BSE 12 pt, BDS 24pt, BCS 24pt 5: References - none 6: Special Prerequisites - none +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ NEW FOR 2004 1: Image Deblocking using Local Segmentation 2: Peter TISCHER 3: The baseline JPEG lossy compression mode has been hugely successful and is widely used, particularly on the internet. The same approach is also used in coding digital video such as videoCD, MPEG3, video-teleconferencing and digital television. However, at low bit rates the block approach leads to objectionable artifacts in the reconstructed image. Local segmentation is a paradigm for constructing image processing algorithms in which the pixels within a block are either classified as belong to the same segment or as belong to a number of different segments. Operations on a pixel will only use values from those neighbouring pixels which belong to the same local segment. As one example, local segmentation has been shown to be effective in creating denoising algorithms for removing Gaussian noise while preserving underlying image structure. The aim of the project is to use knowledge of how the image was compressed in conjunction with local segmentation to produce a reconstructed image with minimal block artifacts. 4: BSE 12 pt, BDS 24 pt, BCS 24 pt 5: References - none 6: Special Prerequisites - none +-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+-+ NEW FOR 2004 1: Document Similarity Measurement via Data Compression 2: Peter TISCHER 3: In document classification it is often desirable to measure how similar two documents are. This might be because we are interested in saying whether documents are spam or whether the source code of two programs is so similar that at least program is an example of plagiarism. If two documents are similar, we would expect that we would require little information to know the contents of the second document given that we already know the contents of the first. One way to measure information is by using data compression techniques as the compressed size is an upper bound on the amount of information needed to store the original data. The aim of the project is to adapt techniques from the area of text compression to compress the data which shows how one document can be created from another. Of particular interest is coming up with similarity measures that are also distance functions. 4: BSE 12 pt, BCS 24 pt 5: References - none 6: Special Prerequisites - none