This paper presents experiments using Radial Basis Function (RBF) networks to tackle the unconstrained face recognition problem using low resolution video information. Input representations that mimic the effects of receptive field functions found at various stages of the human vision system were used with RBF networks that learnt to classify and generalise over different views of each person to be recognised. In particular, Difference of Gaussian (DoG) filtering and Gabor wavelet analysis are compared for face recognition from an image sequence. RBF techniques are shown to provide excellent levels of performance where the view varies and we discuss how to relax constraints on data capture and improve preprocessing to obtain an effective scheme for real-time, unconstrained face recognition. 1. Introduction A face recognition system must be robust with respect to the immense variability of the human face and generalise over a wide range of conditions to capture the essential similariti...

Preprocessing of face images was performed to mimic the effects of receptive field functions found at various stages of the human vision system. These were then used as input representations to Radial Basis Function (RBF) networks that learnt to classify and generalise over different views for a standard face recognition task. Two main organisations of the RBF networks (standard and face unit) and two main types of preprocessing (Difference of Gaussian filtering and Gabor wavelet analysis) were compared. Quantitative and qualitative differences in these schemes are described and conclusions drawn about the best approach for our face recognition problem using low resolution images. 1 Introduction Face recognition has been the subject of a great deal of research in computer vision and work on biologically-motivated approaches has begun to deliver real solutions. One of the main problems is dimensionality reduction to remove much of the redundant information in the original images. There...

This work presents explorations in the microstructure of natural vision systems based on large scale computer simulations. Similarly to previous work in this area, we compute the functional inner products of a two-dimensional input signal (image) with a set of two-dimensional Gabor functions which have been shown to fit the receptive fields of simple cells in the primary visual cortex of mammals. These inner products are then considered as net inputs to the cortical cells and used to compute the cell activations as non-linear functions. A previously used model is extended with a pixel-wise winner-takes-all competition between different Gabor filters which is introduced in order to model lateral inhibition between cortical cells. The effect of lateral inhibition is qualitatively estimated by visualization of computed cortical images and quantitatively evaluated by applying the model to a face recognition problem. Recognition rate of 97% was achieved on a database of 205 face i...

This paper presents experiments using an adaptive learning component based on Radial Basis Function (RBF) networks to tackle the unconstrained face recognition problem using low resolution video information. Firstly, we performed preprocessing of face images to mimic the effects of receptive field functions found at various stages of the human vision system. These were then used as input representations to RBF networks that learnt to classify and generalise over different views for a standard face recognition task. Two main types of preprocessing (Difference of Gaussian filtering and Gabor wavelet analysis) are compared. Secondly we provide an alternative, `face unit' RBF network model that is suitable for large-scale implementations by decomposition of the network, which avoids the unmanagability of neural networks above a certain size. Finally, we show the 2-D shift, scale and y-axis rotation invariance properties of the standard RBF network. Quantitative and qualitativ...

We address the problem of trajectory prediction in machine vision applications using variants of Elman's partially recurrent networks. We use dynamic context to constrain the representation learnt by a network and explore the characteristics of various input representations. Network stability and generalisation from training on complex 2D trajectories are tested. We train such networks to encode knowledge about "trajectories" in dynamic face recognition using an extended "temporal signature" eigenface representation of face image sequences. Eigenvector decomposition on each time step of a motion sequence allows for natural variations in view and scale. This application makes use of on-line head detection and face tracking from image sequences and achieves a high success rate when tested on sequences of known and unknown individuals with large viewpoint di#erences. 1. Introduction Trajectory prediction is an important capability for many computer vision applications, e.g. visual survei...

This paper describes a novel approach to solving two problems inherant in neural networks. The `face unit' network system avoids the unmanagability of neural networks above a certain size by using small, individual networks for each class, and allows the addition of new data to the database without complete re-training of the system.

It is shown how cortical filters can be used for image analysis and object recognition. Similarly to previous work in this area, we compute functional inner products of a twodimensional input signal (image) with a set of two-dimensional Gabor functions which fit the receptive fields of simple cells in the primary visual cortex of mammals. We propose a method in which these inner products become the subject of thresholding, orientation competition and lateral inhibition. Each of the resulting cortical images contains only edge lines of a particular orientation and a particular light-to-dark transition direction. In this way, the information which is present in the original image is split in different channels and we show how this splitting can be used for object recognition. The method discriminates between simple geometrical figures, e.g. polygons with different numbers of edges, with reliability of 100% and a recognition rate of 99% has been achieved when the method was applied to a l...

: We propose to compute a vector field, similar to the optical flow computed between successive frames of an image sequence, which maps optimally (in a certain sense) one face image onto another face image. A cost of the mapping is computed and used to quantify the dissimilarity between the two images. The technique is applied to the problem of person identification by comparing an input face image to all face images prestored in a database. The method was implemented on the Connection Machine CM-5 of the University of Groningen 1 in the data-parallel programming model. Keywords: Optical flow, multiscale resolution, person identification, data parallel programming, Connection Machine CM-5 1. Introduction Person identification by a face image is a problem that has been considered to be a challenge since the very first days of computer vision. One of the first approaches to this problem was based on geometric features, such as size and relative positions of eyes, mouth, nose and chin...

: A computational model of so-called grating cells is proposed. These cells, found in areas V1 and V2 of the visual cortex of monkeys, respond strongly to bar gratings of a given orientation and periodicity but very weakly or not at all to single bars. This non-linear behavior is quite different from the spatial frequency filtering behavior exhibited by the other types of orientation selective cells. It is incorporated in the proposed model by using an AND-like nonlinearity to combine the responses of simple cells and compute the activities of so-called grating subunits which are subsequently summed up. The parameters of the model are adjusted to reproduce the results measured by neurophysiologists with different visual stimuli. The proposed computational model of a grating cell is used to compute the collective activation of sets of such cells, referred to as cortical images, induced by natural visual stimuli. On the basis of the results of such simulations we speculate about the poss...

Face recognition has established itself as an important subbranch of pattern recognition within the field of computer science. Many state-of-the-art systems have focused on the task of recognizing frontal views or images with just slight variations in head pose and facial expression of people. In this work we concentrate on two approaches to recognize profile views (90 degrees) with previous knowledge of only the frontal view, which is a challenging task even for human beings. The first presented system makes use of synthesized profile views and the second one uses a joint parameter estimation technique.