H. Talhami

Three-dimensional pose from two-dimensional images: a novel approach using synergetic networks

Neural networks have been successfully applied in many applications of machine vision. In this work, a synergetic network is used to estimate the pose of a rigid three-dimensional object. The estimation is based on a number of two-dimensional snapshots of the object with known pose. The algorithm at the base of the synergetic computer can be realised as a neural network with a two-layer topology and units that calculate dot products. In the process of constructing this network, the dimensionality of the problem is reduced dramatically from N, the number of pixels, to M, the number of prototype images. In contrast to traditional pose estimation techniques, this approach is based on appearance, rather than a detailed knowledge of shape and reflectance properties, making it flexible and amenable to situations where a detailed description of the object is not available. The algorithm is demonstrated to have fast recall times, opening the possibility of developing a real-time pose estimation system for use with robotic manipulation.

An improved synergetic algorithm for image classification

Abstract A major application of pattern recognition technology is in industrial manufacturing. In this paper, we develop a synergetic algorithm for pattern recognition which is based purely on the appearance of the object, without reference to a CAD model of the object, making the technique generic and flexible. In particular, we apply this algorithm to the problem of classifying an object into a number of user-defined aspects, which is an important problem in robotic manipulation of objects. The technique is fast and can be trained using a non-iterative, deterministic training scheme which will find a zero-error solution on a training set, if such a solution exists.

Model-based assessment of lung structures: inferencing and control system

A general methodology has been developed for computer interpretation of medical images, based on an explicit anatomical model. A test system for analyzing posterior- anterior (PA) chest x-rays has been implemented. The inferencing and control system identifies the major lung structures in the image, and then flags any suspected abnormalities. Image and model data are transformed into a feature space where they are represented in terms of edge descriptions. The inference engine compares the image and model in feature space to label the edges anatomically, and check for normality. The control system schedules events within the inference engine and coordinates interaction with the model and image processing routines. The control architecture is blackboard-based, with a separate data frame for each structure to be identified. The anatomical model uses fuzzy sets to provide ranges of feature values which are considered normal or indicative of a particular abnormality. This allows the inference engine to give a confidence score and linguistic description to each decision. Mediastinum, cardiac border, domes of the diaphragm, ribs and lung outline have been modeled. Their automatic identification allows diagnostic checks such as the cardiothoracic ratio, comparison of right and left lungs to identify lobular collapse and inspection of interfaces in terms of shape and clarity. The inference engine provides simple comments on its findings, making it suitable for pre- and double-checking of images.

Learning in a self-organising pattern formation system

Abstract In this paper we implement learning of a set of unlabelled images using the analogy between pattern formation and pattern learning. Our algorithm is clearer, more robust, and of lower dimension than comparable synergetic algorithms in the literature.

An application of the wavelet transform for the echo signal from a rotating object

The backscattering field from a rotating-moving object is a complex-modulated signal, where the overall scattered field from the object changes over time in both amplitude and phase. Consequently, the Doppler spectrum consists of a central Doppler spectral line caused by the motion and many additional Doppler shift spectral lines caused by the rotation of the object. This paper presents an application of wavelet analysis for the echo signal from a spinning rotor in the time-scale domain. A comparison was also made of the measured-data with the simulated-signal for two rotor speeds through the scalograms of the wavelet transform.

L-Systems for Three-Dimensional Anatomical Modelling: Towards a Virtual Laboratory in Anatomy

L-systems are explored as a tool for the modelling of both branching and non-branching anatomical structures. Branching anatomical structures are modelled using parametric L-systems that are normally used to describe tree growth and structure. Non-branching structures are modelled using generalised cylinder L-systems. Both representations produce highly parametric models of the anatomy that can be used for efficient image generation and model matching. Examples are given to illustrate the power of this approach in modelling three-dimensional anatomy such as the ribcage, the lung airway, and the heart.

Artificial neural networks for force and power predictions in oblique cutting

The importance of oblique cutting as a representative for many practical machining operations is discussed. A few of the existing oblique cutting models and their deficiencies are discussed from a predictive point of view. A neural network architecture is developed to predict the forces and power in single edged oblique cutting operation. Experiments are carried out over a comprehensive range of cutting conditions to verify the predictive capability of the neural network model. The force prediction model using neural network is extensively tested and compared with experimental results using statistical routines.

Explicit inversion: an approach to image analysis

Abstract Image analysis can be expressed as an inverse problem. Given an image, which is the output of some complicated and possibly unknown function, our goal is to estimate the parameters of that function. Formally, at least, the solution to the problem can be found by inverting the function which produced the image. In practice, this inversion requires two major elements; a feature extractor and a parameter estimator. While there has been much research into these two elements, they are generally designed separately from one another. In this paper we introduce an approach to image analysis founded on the belief that these two elements should be designed as a pair. We label our approach ‘explicit inversion’, because it allows us to replace the problem of implicitly inverting an unknown, possibly high-dimensional function, with that of explicitly inverting a known, low-dimensional function. As a result we achieve major time reductions over the standard approaches while achieving comparable accuracy.

Detection of the local maxima in time-frequency and time-scale representations

There is always a trade-off between time and frequency resolutions and crossterms in the transform space even for time-frequency or time-scale representations in signal processing techniques. These problems may strongly affect the detection and recognition of local characteristics of signals. In particular, they become more severe in the case of signals contaminated by noise, and in multiple component signals. An algorithm is presented to extract the local modulus maxima of energy distributions for both time-frequency and time-scale representations. The results show clearly the local features of the investigated signal embedded in noise.

Visualizing ultrasonic vector velocity using the motion without movement algorithm

The visualization of ultrasonic vector velocity data is a non-trivial task. The data is five-dimensional: x, y, velocity magnitude, velocity direction, and time. Most techniques have so far failed to deliver an intuitive approach to displaying such complex data. The motion without movement technique uses time variations in phase to create the illusion of motion without moving the pattern to be displayed. It has been applied to (a) a stationary image, (b) phantom ultrasonic data, and (c) in vivo data and has proved to be a viable alternative to existing techniques. For this study, the speckle projection technique has been used to obtain vector velocity data from a sequence of ultrasonic frames.