Professor Fred Stentiford

Electronic and Electrical Engineering Department,
University College London

Information.


Address :	Department of Electronic & Electrical Engineering
	Torrington Place London WC1E 7JE
Phone :	+44 (0) 1394 411469
Mobile :	+44 (0) 7761 828300
Fax :	+44 (0) 1394 411469
E-mail :

Issues in Pattern Recognition

Pattern Recognition is in a crisis. Performance in recent years has been attributable only to increasing computer power and knowledge about specific applications. Serious questions lie in the following tricky areas:

Systems which rely upon the pre-selection of feature measurements suffer a huge risk of failure on unseen data. The commitment to a fixed set of features constrains performance to a restricted set of solutions for which those features are appropriate.
Representative data is often gathered for training purposes and system optimisation. Almost by definition such data cannot represent unseen data and therefore cannot be expected to compensate for a restricted feature set.
The labelling of data for real life problems is difficult. There is no guarantee that different people will label data in the same way or what effect different labelling will have on system design and ultimate performance.
There is no guarantee that computer search for an optimal recognition performance is ever the best in a real problem unless the probnlem is small enough to allow an exhaustive search. The intuitive heuristic rules of inference used in the search almost always preclude large volumes of the solution space from being entered. The existence of such solutions cannot be proved as this would lead to a contradiction, but equally it cannot be disproved.

The critical and essential element of any recognition system is the ability to make use of what structure is in common between two patterns. Classes of patterns are formed by those groups of patterns which all possess some common structure with others in the group. The structures in common may not be the same throughout the group, but there will be sufficient commonality to characterise the strength of membership of the group. Intuition does not seem to be sufficient to identify such structures, but they are "clearly" present.

Interest Points Reflecting Structure

There are many ways of extracting interest points in images and using them to recognise objects. Rather than define what we think we mean by an object it is of interest to explore what is actually reflected between pairs of images that we believe exhibit similarity. Some work has been done using interest points to show how structure found to be common in pairs of similar images may be used to identify those images amongst others that are not thought to be similar. This brings with it the advantage that it is no longer necessary to have knowledge of patterns belonging to other classes and no training is therefore required save that of analysing certain patterns known to be similar.

More recently it has been shown that this same approach to similarity fails in the same way has human vision when applied to the Poggendorff illusion (Stentiford, 2012).

Here are two images of the same movie poster but recorded under very different conditions. A subset of interest points has been extracted from each image in the b colour channel that possesses the same local gradients and interest points bear the same angular relationship to each other. This type of structure is very unlikely to be found in a different image. Other variations in this image will elicit different subsets of interest points, but each represents a reflection of one image into the other.

Focusing

Focusing and visual accommodation is very much related to attention. Experiments have shown that when global attention measures are maximised across focal planes, the principal subject becomes in focus.

Brief CV

I studied mathematics at St Catharine's College, Cambridge, and obtained a PhD in Pattern Recognition at Southampton University. I first joined the Plessey Company to work on various applications including the recognition of fingerprints and patterns in time varying magnetic fields. I then joined BT and carried out research on optical character recognition and speech recognition. After that I led a team developing systems employing pattern recognition methods for the machine translation of text and speech. This work led to the world's first demonstration of automatic translation of speech between different languages.

I then moved into designing dialogues for new telephone services and managed the government funded collaborative Dialogues 2000 project which aimed to research and promote common standards in the spoken user interface in UK industry. The membership of over 200 companies was a measure of its success.

I returned to vision research to lead a group developing new algorithms for analysing and delivering multimedia content and more recently joined UCL to pursue this research more intensively.

I am a corporate member of the IEE and the BCS.

I look after the Boyton village website.

Research Interests

My current interests are in the field of Pattern Recognition and Machine Vision. I have an open mind on the usefulness of mathematical theory in this area of research preferring to rely upon experiment to determine the direction that investigation should take. I have a special interest in evolution and why it works.

Specific topics of interest include:

Visual attention

Similarity measures

Content-Based Image Retrieval

HCI