Clustering data into groups so that each group is characterised by a set of unique features is of great significance in many applications. Although this is a well-studied area, a fundamental problem remains unsolved: clustering databases with the Number of Clusters far greater than the average Size of Clusters (or “NC >> SC” problem for short). The technical challenge of this problem arises from the fact that the intra-cluster similarity of some clusters is likely to be greater than the inter-cluster similarity of some other clusters. As a result, in the feature space, the clusters significantly overlap, making the clustering unreliable. A typical example is INTERPOL ‘s International Child Sexual Exploitation Database (ICSE DB) with millions of photos. Because often the metadata (e.g., camera model) of the photos is manipulated, the police expect that there may be hundreds of thousands of cameras responsible for the creation of those photos, each responsible for only a small number of photos. The objective of this project is to develop a heuristic clustering method to cluster those photos so that each cluster corresponds to one camera.

Hand gesture recognition (HGR) is a fast emerging technology (FET) for facilitating Human-Computer Interaction (HCI). The past few years have seen attempts to exploit its potential in applications such as gaming, sign language recognition, contactless medical document browsing, assisted living and virtual fitting room. However, at the current infant stage of this FET, the common limitations of the existing hand gesture recognition techniques are 1) rigid constraints: the hand movement of a complete gesture has to be in view of the camera, 2) high ambiguity: the similarity between sub-gestures of different gestures may be so high that confusions occur frequently, 3) high computational complexity, and 4) high sensitivity to foregournd and background human interference. It is our intention in this project to study the feasibility of a new efficient hand gesture recognition technique that is able to alleviate those limitations.

Digital cameras, camcorders and mobile phones rely on semiconductor sensor to capture images. Because the semiconductor wafers are not perfectly made and the imperfection is unique, the imperfection gives rise to the so-called Sensor Pattern Noise (SPN) in the images. Even devices of the same model have unique SPN, which can be used as device “fingerprints”. Functional Technologies Ltd has developed a Forensic Image Analyser (FIA) to exploit SPN for forensic applications. However, the SPN is as big as the original image. The high computational complexity due to the size of the SPN has limited the FIA's applicability. This project aims at studying the feasibility of using a Compact Attribute SEt (CASE) to represent the SPN. If successful, the computational cost can be reduced by hundreds of thousands of times.

Awaiting Public Summary