MOTIVATION: Probabilistic latent semantic analysis (pLSA) is commonly applied to describe mass spectra (MS) images. However, the method does not provide certain outputs necessary for the quantitative scientific interpretation of data. In particular, it lacks assessment of statistical uncertainty and the ability to perform hypothesis testing. We show how linear Poisson modelling advances pLSA, giving covariances on model parameters and supporting χ2 testing for the presence/absence of MS signal components. As an example, this is useful for the identification of pathology in MALDI biological samples. We also show potential wider applicability, beyond MS, using magnetic resonance imaging (MRI) data from colorectal xenograft models. RESULTS: Simulations and MALDI spectra of a stroke-damaged rat brain show MS signals from pathological tissue can be quantified. MRI diffusion data of control and radiotherapy-treated tumours further show high sensitivity hypothesis testing for treatment effects. Successful χ2 and degrees-of-freedom are computed, allowing null-hypothesis thresholding at high levels of confidence. AVAILABILITY AND IMPLEMENTATION: Open-source image analysis software available from TINA Vision, www.tina-vision.net. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Motivation: Matrix-assisted laser desorption/ionisation time-of-flight mass spectrometry (MALDI) facilitates the analysis of large organic molecules. However, the complexity of biological samples and MALDI data acquisition leads to high levels of variation, making reliable quantification of samples difficult. We present a new analysis approach that we believe is well-suited to the properties of MALDI mass spectra, based upon an Independent Component Analysis derived for Poisson sampled data. Simple analyses have been limited to studying small numbers of mass peaks, via peak ratios, which is known to be inefficient. Conventional PCA and ICA methods have also been applied, which extract correlations between any number of peaks, but we argue makes inappropriate assumptions regarding data noise, i.e. uniform and Gaussian. Results: We provide evidence that the Gaussian assumption is incorrect, motivating the need for our Poisson approach. The method is demonstrated by making proportion measurements from lipid-rich binary mixtures of lamb brain and liver, and also goat and cow milk. These allow our measurements and error predictions to be compared to ground truth. Availability and implementation: Software is available via the open source image analysis system TINA Vision, www.tina-vision.net. Contact: [email protected]. Supplementary information: Supplementary data are available at Bioinformatics online.

Motivation: Imaging demonstrates that preclinical and human tumors are heterogeneous, i.e. a single tumor can exhibit multiple regions that behave differently during both development and also in response to treatment. The large variations observed in control group, tumors can obscure detection of significant therapeutic effects due to the ambiguity in attributing causes of change. This can hinder development of effective therapies due to limitations in experimental design rather than due to therapeutic failure. An improved method to model biological variation and heterogeneity in imaging signals is described. Specifically, linear Poisson modeling (LPM) evaluates changes in apparent diffusion co-efficient between baseline and 72 h after radiotherapy, in two xenograft models of colorectal cancer. The statistical significance of measured changes is compared to those attainable using a conventional t-test analysis on basic apparent diffusion co-efficient distribution parameters. Results: When LPMs were applied to treated tumors, the LPMs detected highly significant changes. The analyses were significant for all tumors, equating to a gain in power of 4-fold (i.e. equivalent to having a sample size 16 times larger), compared with the conventional approach. In contrast, highly significant changes are only detected at a cohort level using t-tests, restricting their potential use within personalized medicine and increasing the number of animals required during testing. Furthermore, LPM enabled the relative volumes of responding and non-responding tissue to be estimated for each xenograft model. Leave-one-out analysis of the treated xenografts provided quality control and identified potential outliers, raising confidence in LPM data at clinically relevant sample sizes. Availability and implementation: TINA Vision open source software is available from www.tina-vision.net. Supplementary information: Supplementary data are available at Bioinformatics online.

This paper describes ongoing work on the segmentation of the optic disc in retinal images using pixel classification and circular template matching. The pixel property used for classification is based on texture. Two texture measurements were used, Binary Robust Independent Elementary Features (BRIEF) and a rotation invariant BRIEF (OBRIEF). This texture measurement is chosen because it can address the illumination issues of the retinal images and has a lower degree of computational complexity than most of the existing texture measurement methods currently used in the literature. The method was tested on a set of 196 images composed of 110 healthy retina images and 86 glaucomatous images. The average mean overlapping ratio between the true optic disc region and the segmented region is 0.81 and 0.82 for BRIEF and OBRIEF respectively. Comparison with a method based on the Hough Transform and Snake is also provided.

We propose a shape analysis system based upon the description of landmarks with measurement covariance, which extends statistical linear modelling processes to 'pseudo landmarks' for scientific studies.We discuss the properties of our approach and how measurement covariances can be considered characteristic of the local shape.Our formulation includes corrections for parameter bias, induced by the degrees of freedom within the linear model.The method has been implemented and tested on measurements from fly wing, hand and face data.We use these data to explore possible advantages and disadvantages over the use of standard Procrustes/PCA analysis.In the process we show how appropriate weighting provides more efficient use of the available landmark data 1 .

Our aim is to develop a completely automated and reliable system to identify morphological landmarks in digital images. The performance of the system is aimed to replicate manual digitization with equiv-alent accuracy and reliability, based upon a small number of training examples. The analysis system is constructed from four stages; a feature based detection of fly wing structure, correspondence matching based upon the pairwise geometric histogram (PGH) representation, global location of the wing using a Probabilistic Hough Transform (PHT), and finally local correlation based refinement of individual features. We evaluate this system and compare quantitative results to manually digitized data. 1

This paper applies statistical design principles to a simple biological model of human vision so that we can more clearly interpret the apparent role of eye saccades. In doing so we show that many structural features of the biological system (such as the optical geometry of the retina) are strategies for minimising the resources required to construct a working image recognition system. The ideas presented have implications for the construction of arti-ficial (computer) vision systems. The computational model is very closely related to (but not based upon) SIFT, but more strongly based on a consideration of vision as a process of measurment while also linking the idea of multi-scale analysis with biological structure.

In this paper, we present a tracking framework for capturing articulated human motions in real-time, without the need for attaching markers onto the subject's body. This is achieved by first obtaining a low dimensional representation of the training motion data, using a nonlinear dimensionality reduction technique called back-constrained GPLVM. A prior dynamics model is then learnt from this low dimensional representation by partitioning the motion sequences into elementary movements using an unsupervised EM clustering algorithm. The temporal dependencies between these elementary movements are efficiently captured by a Variable Length Markov Model. The learnt dynamics model is used to bias the propagation of candidate pose feature vectors in the low dimensional space. By combining this with an efficient volumetric reconstruction algorithm, our framework can quickly evaluate each candidate pose against image evidence captured from multiple views. We present results that show our system can accurately track complex structured activities such as ballet dancing in real-time.

One of the most common problems in image analysis is the estimation and removal of noise or other artefacts using spatial filters. Common techniques include Gaussian, Median and Anisotropic Filtering. Though these techniques are quite common they must be used with great care on medical data, as it is very easy to introduce artifact into images due to spatial smoothing. The use of such techniques is further restricted by the absence of a ‘gold standard ’ data against which to test the behaviour of the filters. Following a general discussion of the equivalence of filtering techniques to likelihood based estimation using an assumed model, this paper describes an approach to noise filtering in multi-dimensional data using a partial volume data density model. The resulting data sets can then be taken as a gold standard for spatial filtering techniques which use the information from single images. We demonstrate equivalence between the results from this analysis and techniques for performance characterisation which do not require a ‘gold standard’. 1

For image analysis techniques to be of utility in medical diagnosis systems it is necessary to be able to perform quality control over the results they produce. Input data must conform to the assumptions within the algorithm if useful results are to be achieved. Automation of this process is essential if vision algorithms are to form components in analysis systems. In this paper we present a technique to validate the correction of field inhomogeneity in MR images. The initial intention was to use information measures to check the improvement due to correction. However, it will be shown that the standard log entropy calculation for information measurement does not have the required properties, specifically grey-level scale invariance. We present an alternative, scale-invariant information measure derived using conventional likelihood approaches, that can be applied as an absolute measure of information content. We show this technique in use for the validation of our existing coil correction method.

We describe the development of a vision system to detect natural events in a low-resolution image stream. The work involves the assessment of algorithmic design decisions to maximise detection reliability. This assessment is carried out by comparing measures and estimates made by the system under development with measures obtained independently. We show that even when these independent measures are themselves noisy, their independence can serve to guide design decisions and allow performance estimates to be made. Although presented here for one particular system design, we believe that such an approach will be applicable to other situations when an image-based system is to be used in the analysis of natural scenes in situations where a precise ground truth is not available.

A novel colour image segmentation routine, based on clustering pixels in colour space using non-parametric density estimation, is described. Although the basic methodology is well known, several important improvements to the previous work in this area are introduced. The density is estimated at a series of knot points in the colour space, and clustering is performed by hill climbing on this density function. The hill climbing is constrained such that no step crosses an intermediate Voronoid cell, ensuring that all salient clusters are detected. Most importantly, the problem of scale selection has been addressed using a statistically motivated approach, by placing the knot points according to an estimate of the noise in the original images, taking full account of error propagation in the algorithm. The algorithm has been evaluated both on synthetic data and in the context of its application in a machine vision system, specifically the calibration of velocity estimates extracted from a novel infrared sensor used in a fall detector. The application of the technique to medical images and texture recognition is also discussed.

This paper discusses advances made to the 3D geometrical model matching system within the TINA machine vision environment over the last 10 years including the recent inclusion of a verification stage. We show how this final step closes the loop on the object location system allowing the theoretical location performance to be attained, eliminating key assumptions used in the existing forward pass algorithms. We explain how the use of a system approach has been crucial to the development of key components and summarise our findings.

Image subtraction is used in many areas of machine vision to identify small changes between equivalent pairs of images. Often only a small subset of the differences will be of interest. In motion analysis only those differences caused by motion are important, and differences due to other sources only serve to complicate interpretation. Simple image subtraction detects all differences regardless of their source, and is therefore problematic to use. Superior techniques, analogous to standard statistical tests, can isolate localised differences due to motion from global differences due, for example, to illumination changes. Four such techniques are described. In particular, we introduce a new non-parametric statistical measure which allows a direct probabilistic interpretation of image differences. We expect this to be applicable to a wide range of image formation processes. Its application to medical images is discussed.

This paper compares the use of RANSAC for the determination of epipolar geometry for calibrated stereo reconstruction of 3D data with more conventional optimisation schemes. The paper illustrates the poor convergence efficiency of RANSAC which is explained by a theoretical relationship describing its dependency upon the number of model parameters. The need for an a-priori estimate of outlier contamination proportion is also highlighted. A new algorithm is suggested which attempts to make better use of the solutions found during the RANSAC search while giving a convergence criteria which is independent of outlier proportion. Although no significant benefit can be found for the use of RANSAC on the problem of stereo camera calibration estimation. The new algorithm suggests a simple way of improving the efficiency of RANSAC searches which we believe would be of value in a wide range of machine vision problems.

Temporal stereo vision algorithms can offer improved robustness, however, this can only be delivered after several frames of a stereo image sequence have been processed. We present a new method of bootstrapping temporal stereo which can overcome such start-up problems by applying additional coarse-to-fine pre-processing to the first few images in a stereo sequence. To gauge the performance of temporal bootstrapping, we have employed a new algorithmic evaluation technique that uses statistical and physical scene modelling to produce accurate result errors data. The performance of the bootstrap temporal stereo algorithm, as determined by the automatic evaluation technique, as well the results from real stereo image sequences, are presented. 1

We explain how most conventional approaches to the problem of visually guided robot navigation are fundamentally restricted. Starting with a specification of the requirements of a general purpose visually guided robot which includes, robustness, autonomy, flexibility and scalability, we then develop an approach to the solution of map building and path planning which operates within a learning framework. We suggest a new representation of shape using features based on geometric co-occurrence which has the capabilities of providing such metric information and is expected to have similar recognition properties to other techniques based on relational geometric measures. 1

We are currently involved in an industrial project to recover depth information from stereo image pairs retrieved using a scanning electron microscope, (SEM). Feature based approaches to stereo provide accurate disparity estimations, however the quantity of estimates recovered is small (typically 1-2% of the image). If a continuous approximation to the surface is to be reconstructed, as requested by potential customers, more data has to be recovered. Our approach involves using the disparity estimates from a feature based stereo algorithm to constrain a function fitting process. Assuming the image may be represented by an iterated facet model, the algorithm attempts to fit piecewise polynomials between the feature disparity estimates, which describe the mapping of grey-levels from the left to right image along epi-polars. The problems of illumination variation between the left and right images have been addressed using a modification to rank-order filtering which we call ...

The problem of model selection is endemic in the machine vision literature, yet largely unsolved in the statistical field. Our recent work on a theoretical statistical evaluation of the Bhattacharyya similarity metric has led us to conclude that this measure can be used to provide a solution. Here we describe how the approach may be extended to solve the problem of model selection during the functional fitting process. This paper outlines the motivation for this work and a preliminary study of the use of the technique for function fitting. It is shown how the application of the method to polynomial interpolation provides some interesting insights into the behaviour of these statistical methods and suggestions are given for possible uses of this technique in vision. 1 System Identification In some areas of machine vision the correct model required to describe a system can be derived without any real ambiguity. These situations are exemplified by the process of camera calibration where a good understanding of the physical processes involved allow us to define models which encompass the main optical and electronic effects. In other cases, however, selecting the correct model to describe a set of data is something that cannot be specified uniquely a-priori. Complicated models result in reduced ability to make predictions, but simple models

This paper describes a novel approach for visual speech recognition. The shape of the mouth is modelled by an active shape model which is derived from the statistics of a training set and used to locate, track and parameterise the speaker's lip movements. The extracted parameters representing the lip shape are modelled as continuous probability distributions and their temporal dependencies are modelled by hidden Markov models. We present recognition tests performed on a database of a broad variety of speakers and illumination conditions. The system achieved an accuracy of 85.42% for a speaker independent recognition task of the first four digits using lip shape information only.

Implementing electronic systems in a low-cost, three-dimensional multichip module (MCM) technology provides significantly improved system density over planar packaging technologies. However, the extra dimension necessitates a new approach to design automation. Through experience gained in manually partitioning electronic systems, generic design rules have been formulated and implemented in software. The approach adopted is based on a genetic algorithm which can produce multiple candidate solutions and has the potential to simultaneously optimize multiple design criteria. This paper presents the initial results for the automatic configuration of a variety of systems. The results obtained by the algorithm are compared to known optimal solutions for "test" cases and to the results obtained by manual partitioning for the first heterogeneous MCM-V system.

This paper describes a robust method for extracting visual speech information from the shape of lips to be used for an automatic speechreading (lipreading) systems. Lip deformation is modelled by a statistically based deformable contour model which learns typical lip deformation from a training set. The main difficulty in locating and tracking lips consists of finding dominant image features for representing the lip contours. We describe the use of a statistical profile model which learns dominant image features from a training set. The model captures global intensity variation due to different illumination and different skin reflectance as well as intensity changes at the inner lip contour due to mouth opening and visibility of teeth and tongue. The method is validated for locating and tracking lip movements on a database of a broad variety of speakers.

An approach for person identification is described based on spatio-temporal analysis of the talking face. A person is represented by a parametric model of the visible speech articulators and their temporal characteristics during speech production. The model consists of shape parameters, representing the lip contour and intensity parameters representing the grey level distribution in the mouth region. The model is used to track lips in image sequences where the model parameters are recovered from the tracking results. While some of these parameters relate to speech information, others are intuitively related to different persons and we show that models based on these features enable successful person identification. We model the shape and intensity parameters as mixtures of Gaussians and their temporal dependencies by hidden Markov models. Identifying a talking person is performed by estimating the likelihood of each model for having generated the observed sequence of features and the model with the highest likelihood is chosen as the identified person.

We describe a speechreading system that uses both, shape information from the lip contours and intensity information from the mouth area. Shape information is obtained by tracking and parameterising the inner and outer lip boundary in an image sequence. Intensity information is extracted from a grey level model, based on principal component analysis. In comparison to other approaches, the intensity area deforms with the shape model to ensure that similar object features are represented after non-rigid deformation of the lips. We describe speaker independent recognition experiments based on these features and Hidden Markov Models. Preliminary results suggest that similar performance can be achieved by using either shape or intensity information and slightly higher performance by their combined use.

This paper describes a new approach for speaker identification based on lipreading. Visual features are extracted from image sequences of the talking face and consist of shape parameters which describe the lip boundary and intensity parameters which describe the grey-level distribution of the mouth area. Intensity information is based on principal component analysis using eigenspaces which deform with the shape model. The extracted parameters account for both, speech dependent and speaker dependent information. We built spatio-temporal speaker models based on these features, using HMMs with mixtures of Gaussians. Promising results were obtained for text dependent and text independent speaker identification tests performed on a small video database.

The recognition of shapes in images using Pairwise Geometric Histograms has previously been confined to fixed scale shape. Although the geometric representation used in this algorithm is not scale invariant, the stable behaviour of the similarity metric as shapes are scaled enables the method to be extended to the recognition of shapes over a range of scale. In this paper the necessary additions to the existing algorithm are described and the technique is demonstrated on real image data. Hypotheses generated by matching scene shape data to models have previously been resolved using the generalised Hough transform. The robustness of this method can be attributed to its approximation of maximum likelihood statistics. To further improve the robustness of the recognition algorithm and to improve the accuracy to which an objects location, orientation and scale can be determined the generalised Hough transform has been replaced by the probabilistic Hough transform. 1

The use of image features greatly improves the computational efficiency of subsequent machine vision tasks by directing processing effort to information rich areas of the image. However, the extraction of image features is an intensive process where the usual mapping between image data and feature is only approximately defined. Further, feature enhanced images provide very little information regarding the reliability of the information they embody. This results from the application of heuristics during the later stage of feature extraction. This paper describes an artificial neural network trained to perform feature detection. The network is ‘tuned ’ to extract a particular feature type. This is achieved using the functional mapping of an existing detection algorithm once systematic errors in this technique were removed. The scale of the mapping problem is reduced by enhancing the invariant characteristics of the feature. Only then is a manageable sized network able to perform the mapping task. The network enables the conditional probability of a corner to be used in the final image segmentation. 2

In order to successfully track moving objects it is necessary to understand their motion. Such motion will inevitably change with time, thus attempting to fit the same model all of the time is inappropriate. An evaluation of the single model Kalman filter described in this paper demonstrates this. A maximum model coverage technique, however, would provide a intractable solution because of the exponentially increasing number of models required. We present a solution which uses a finite set of models all making predictions on the data. One model is selected from the set on the basis that it best accounts for the data. A Kalman filter is then used to refine this model whilst it is appropriate. The remaining models continue to make predictions and at any time the filtered model may be replaced by one that more accurately describes the data. Experimental results demonstrate the improvement the switchable model Kalman filter provides over the single model Kalman filter. Problems using the chi-squared metric for model selection are discussed and a more appropriate metric, the Bhattacharyya integral, introduced. The desired properties of this integral are exposed via its analytic solution. It is believed that the statistical methods underlying this work may also have applications in other system identification problems. 1

This paper gives an example of how computer vision algorithms can be reformulated to exploit correlation based hardware without compromising the underlying principles of the algorithm. The work shows results of the stretch-correlation algorithm for calibrated stereo depth estimation and goes on to discuss the development of a convolution chip for implementing this algorithm. The motivation for the chip and its applicability to other computer vision algorithms is also discussed.

A pairwise geometric histogram (PGH) encodes the probability of geometric co-occurrences between any line and the set of lines defining an object. An object therefore has a set of PGHs associated with it, one histogram for each line. We describe here the way in which the probability of geometric co-occurrence is calculated and entered in the histograms, the different ways these histograms can be denned and the completeness properties of the set of histograms in terms of arbitrary shape representation. We show that this representation provides unambiguous shape representation by demonstrating an inverse reconstruction algorithm. We conclude that the methods suggested in a previous paper for object recognition and location provide a complete solution to the problem of recognition of edge based descriptors for fixed 2D projected views of rigid objects. 1

The computational and communication requirements of a wide range of image processing algorithms were analysed and used to specify the design of a general purpose image processing device. The algorithms considered are restricted to image-in and image or feature-out type algorithms. These algorithms are categorised by the demands that they would place on a hardware architecture. We conclude that the major consideration is the off-chip input/output limitation which needs to be able to exploit redundancy in both the data and processor instructions. This paper summarises: the main conclusions of this work, an architecture which we have designed in an attempt to meet the identified constraints and how it differs from previous attempts.

We introduce a novel form of shape representation based on recording the distribution of pairwise geometric relationships between local shape features. It is shown that the geometric histograms used to record these distributions can be easily and robustly acquired from image data and can support recognition even when the shape extracted from the image is badly degraded by fragmentation noise and occlusion. Moreover, the processing involved in establishing correspondences between model and image features is both simple and parallel and has many advantages over previous search based methods. 1

Describes image rectification, explains why it is necessary, how it is performed and proposes a hardware implementation for realtime operation as well as outlining a number of important additional applications that such a module may support, including ground plane obstacle detection.>

Representation of arbitrary shape for the purposes of visual object recognition is an unsolved problem. This paper outlines some theoretical properties of shape representations based on stored geometric constraints that make them suitable for input to a 3D object recognition system. The construction of such representations is detailed together with an evaluation of their efficacy based on experimental findings.>

Abstract Many stereo correspondence algorithms require relative camera geometry, as the epipolar constraint is fundamental to their matching processes. We intend to build a eye/head camera rig to mount on the mobile platform COMODE to enhance the abilities of the TINA system to recover 3D geometry from its environment. Thus we need to be able to associate camera geometry with particular head configurations. Generic calibration of such a system would require the ability to compute camera geometry from arbitrary stereo images. A system which solves this problem using an established corner detector combined with a robust stereo matching algorithm and a variational solution for the camera geometry is described.