Pattern Recognition and Machine Learning Research Group at Letterkenny Institute of Technology

2006

YEAR BOOK

Home Page	Table of contents	Index by Author	Index by topics
Search

Letterkenny Institute of Technology

Christopher Hudy & Jonathan Campbell

Pattern Recognition and Machine Learning Research Group at Letterkenny Institute of Technology

Pattern recognition is the science of enabling computers to see and hear and understand. The PRML research group's current major projects involve identification of shellfish larvae in microscope images and automated analysis of ballot papers (electronic voting). The shellfish larvae project is in collaboration with the LIT Centre for Marine Biology (CAMBIO) and is funded by Higher Education Authority Strand 1.

The electronic voting project is an InterTradeIreland FUSION funded project with OPT2VOTE, the Derry-based software development company specialising in the delivery of e-enabled elections. It is noteworthy that the systems developed by OPT2VOTE for British national elections use normal paper ballot papers, hence an audit trail is assured and manual checking can be carried out at any stage.

Previous projects undertaken by members of the group include inspection of denim fabric, land-use mapping using satellite images, and financial prediction. A number of biomedical imaging projects are in the proposal stage.

The objective of the shellfish recognition work is to help automate the prediction of the optimum time for harvesting shellfish seed; the collection of natural shellfish spat or seed as a source of raw material is a highly valuable shellfish aquaculture industry.

However, due to the time consuming and expertise intensive nature of the task (hours spent on every sample), automated analysis and identification is desirable. In addition, we feel that the techniques being developed will be applicable to a wide range of microscopy work and general machine vision, notably medical applications.

Figure 1. Sample containing mussel and scallop larvae.

The identification starts with an image such as Figure 1. The next stage is segmentation that is, separating the larvae from the background. In many cases segmentation is simply a matter of

Figure 2. (a)Watershed segmentation, (b)Occlusion (c) Edge enhanced.

thresholding: dark = larva, light = background. In this case the problem is made a lot more challenging by the fact that the larvae are translucent and often parts of the larva object are as bright as the background. Generally it is found that edge-detection-based methods are most promising. For example Figure 2(a) shows the result of watershed segmentation; the watershed method works by considering the gradient of image as a topography; typically the edges form a shape something like the rim of a bowl and the numerical equivalent of flooding with water allows the delineation of watersheds, the boundaries outlining objects. Once we have determined the boundary, we can apply standard two-dimensional shape recognition techniques.

In many cases we are presented with occluding background objects or other larvae, Figure 2(b). Here the edges are faint edges, Figure 2(c), and overlapping; hence boundary edges must be tracked.

Figure 3. (a) Tracked boundary, (b) interpolated missing edge points, (c) segmented two-dimensional shape.

Instead of using two-dimensional shape recognition, methods based on boundaries can be used, for example Fourier transforms, where boundaries are summarised using mathematics closely related to those used in MP3 audio coding and JPEG picture compression.

Contact: Jonathan Campbell, Department of Computing, Letterkenny Institute of Technology.
E-mail: [email protected]