Search

3D face recognition is an increasing popular modality for biometric authentication, for example in the iPhoneX. Landmarking plays a significant role in region based face recognition algorithms. The accuracy and consistency of the landmarking will directly determine the effectiveness of feature extraction and hence the overall recognition performance. While surface normals have been shown to provide high performing features for face recognition, their use in landmarking has not been widely explored. To this end, a new 3D facial landmarking algorithm based on thresholded surface normal maps is proposed, which is applicable to widely used 3D face databases. The benefits of employing surface normals are demonstrated for both facial roll and yaw rotation calibration and nasal landmarks localization. Results on the Bosphorus, FRGC and BU-3DFE databases show that the detected landmarks possess high within-class consistency and accuracy under different expressions. For several key landmarks the performance achieved surpasses that of state-of-the-art techniques and is also training free and computationally efficient. The use of surface normals therefore provides a useful representation of the 3D surface and the proposed landmarking algorithm provides an effective approach to localising the key nasal landmarks.

The use of surface metrology microscopes to quantify surface texture is a powerful tool for the analysis of archaeological materials. Data collected from these microscopes allows for reliable and reproducible measurements of surface characteristics. However, archaeological materials provide some unique challenges for microscopic analysis, including archaeological objects that cannot leave museums or are too large to observe under a microscope. Because of these challenges, many researchers create molds and casts of an artifact's surface prior to measuring surface texture. The replicate surface is assumed to be an accurate representation of the original surface texture; however, limited testing has been done on the reliability of different molding compounds. This paper uses confocal microscopy to test the fidelity of different molding compounds for surface texture measurements. Results suggest that the objects' original surface texture can pose challenges to the accuracy of molds and the resulting casts, and that the intrinsic texture of the molding material impacts surface measurements of the replicated object.

In recent decades, many 3D data acquisition methods have been developed to provide accurate and cost-effective 3D captures of the human face. An example system, which can accommodate both research and commercial applications, is the Photoface device. Photoface is based on the photometric stereo imaging technique. To improve the recognition performance using Photoface captures, a novel landmarking algorithm is first proposed by thresholding surface normals maps. The development of landmarking algorithms specifically for photometric stereo captures enables region-based feature extraction and fills a gap in the 3D face landmarking literature. Nasal curves and spherical patches are then used respectively for recognition and are evaluated on the 3DE-VISIR database, which contains Photoface captures with expressions. The neutral vs. non-neutral matching results demonstrate high face recognition performance using spherical patches and a KFA classifier, achieving a R1RR of 97.26% when only 24 patches are selected for matching.

The potential of the nasal region for expression robust 3D face recognition is thoroughly investigated by a novel five-step algorithm. First, the nose tip location is coarsely detected and the face is segmented, aligned and the nasal region cropped. Then, a very accurate and consistent nasal landmarking algorithm detects seven keypoints on the nasal region. In the third step, a feature extraction algorithm based on the surface normals of Gabor-wavelet filtered depth maps is utilised and, then, a set of spherical patches and curves are localised over the nasal region to provide the feature descriptors. The last step applies a genetic algorithm-based feature selector to detect the most stable patches and curves over different facial expressions. The algorithm provides the highest reported nasal region-based recognition ranks on the FRGC, Bosphorus and BU-3DFE datasets. The results are comparable with, and in many cases better than, many state-of-the-art 3D face recognition algorithms, which use the whole facial domain. The proposed method does not rely on sophisticated alignment or denoising steps, is very robust when only one sample per subject is used in the gallery, and does not require a training step for the landmarking algorithm. https://github.com/mehryaragha/NoseBiometrics

The Hit-or-Miss Transform (HMT) is a powerful morphological operation that can be utilised in many digital image analysis problems. Its original binary definition and its extension to grey-level images have seen it applied to various template matching and object detection tasks. However, further extending the transform to incorporate colour or multivariate images is problematic since there is no general or intuitive way of ordering data which allows the formal definition of morphological operations in the traditional manner. In this paper, instead of following the usual strategy for Mathematical Morphology, based on the definition of a total order in the colour space, we propose a transform that relies on a colour or multivariate distance measure. As with the traditional HMT operator, our proposed transform uses two structuring elements (SE) - one for the foreground and one for the background - and retains the idea that a good fitting is obtained when the foreground SE is a close match to the image and the background SE matches the image complement. This allows for both flat and non-flat structuring elements to be used in object detection. Furthermore, the use of ranking operations on the computed distances allows the operator to be robust to noise and partial occlusion of objects.

This study proposes an algorithm is proposed to quantitatively evaluate the performance of three-dimensional (3D) holistic face recognition algorithms when various denoising methods are used. First, a method is proposed to model the noise on the 3D face datasets. The model not only identifies those regions on the face which are sensitive to the noise but can also be used to simulate noise for any given 3D face. Then, by incorporating the noise model in a novel 3D face recognition pipeline, seven different classification and matching methods and six denoising techniques are used to quantify the face recognition algorithms performance for different powers of the noise. The outcome: (i) shows the most reliable parameters for the denoising methods to be used in a 3D face recognition pipeline; (ii) shows which parts of the face are more vulnerable to noise and require further post-processing after data acquisition; and (iii) compares the performance of three different categories of recognition algorithms: training-free matching-based, subspace projection-based and training-based (without projection) classifiers. The results show the high performance of the bootstrap aggregating tree classifiers and median filtering for very high intensity noise. Moreover, when different noisy/denoised samples are used as probes or in the gallery, the matching algorithms significantly outperform the training-based (including the subspace projection) methods.

Morphological hierarchical segmentation of color images may be achieved in a straightforward way by measuring the persistence of regional minima from color gradients and using these measurements as a criterion to select markers for the watershed from markers framework. Since color has an implicit role in the selection of markers, the segmentation process may provide a bad combination of distinct colored regions, and this may lead to a distorted image simplification. This paper proposes a new method to color image simplification in which the importance of color is raised because color information is added to the marker selection process. Such method provides finer control over the final number of regions (n) and the resulting number of colors (c). A color clustering method splits the regional minima in to c minima sets, each of which has a representative color. The most prominent regional minima from each minima set are selected to form the markers for the segmentation framework. In the final segmentation, the color assigned to a region is given by the representative color bound to the marker that points to the region. It leads to an image whose segmented regions are quantized to fewer distinct colors.

Area morphological scale-spaces are widely used for hierarchical image analysis and segmentation. Despite their advantages, their extension to colour images has been restricted by the lack of an explicit order relationship for vector values. This paper presents a theoretical evaluation of two recently proposed colour sieves and their properties. It is also demonstrated that the extrema definition used by a colour sieve determines both the aggressiveness of its sieving action and its processing speed. A new colour sieve structure is introduced that attempts to capture the relative advantages of the two sieves previously studied. An objective study of the noise reduction performance of these colour sieves is presented. The segmentation performance is also analysed using the methodology provided by the Berkeley Segmentation Dataset and Benchmark, both in terms of the overall segmentation performance and its robustness to image noise. The new colour sieve is shown to have the best overall segmentation performance, and to be the most robust.

This paper proposes a novel local depth and surface normals descriptor to explore the discriminative features on the nasal surface and the adjoining cheek regions for expression robust 3D face recognition. After preprocessing the 3D face data, landmarks located on the perimeter of a triangular region covering the nose and adjoining parts of the cheeks are accurately detected. Inspired by Local Binary Patterns, local shape differences for 3D points on a set of horizontal curves joining selected landmarks provide a novel representation of the local shape information. A further analysis of the discriminatory power of each patch shows that the adjoining regions have the potential to produce good recognition performance. Using the FRGC and Bosphorus databases, the performance of theproposed descriptor is evaluated on diverse patches, scales and for four components, one from the depth and three from the surface normals. Results show that the new local shape descriptor performs well at representing the shape information on a relatively large scale. On the basis of this descriptor, a relatively small set of features extracted from the nasal and adjoining cheekregions produce a R1RR of 97.76% and an EER of 1.32%. The adjoining cheek regions demonstrate a high discriminatory power and provide a useful new addition to 3D face biometrics.

This paper investigates the performance of different representations of 3D human nasal region for expression robust recognition. By performing evaluations on the depth and surface normal components of the facial surface, the nasal region is shown to be relatively consistent over various expressions, providing motivation for using the nasal region as a biometric. A new efficient landmarking algorithm that thresholds the local surface normal components is proposed and demonstrated to produce an improved recognition performance for nasal curves from both the depth and surface normal components. The use of the Shape Index for feature extraction is also investigated and shown to produce a good recognition performance.

In the past decade, expression variations have been one of the most challenging sources of variability in 3D face recognition, especially for scenarios where there are a large number of face samples to discriminate between. In this paper, an expression robust reject or is proposed that first robustly locates landmarks on the relatively stable structure of the nose and its environs, termed the cheek/nose region. Then, by defining curves connecting the landmarks, a small set of features (4 curves with only 15 points each) on the cheek/nose surface are selected using the Bosphorus database. The resulting reject or, which can quickly eliminate a large number of candidates at an early stage, is further evaluated on the FRGC database for both the identification and verification scenarios. The classification performance using only 60 points from 4 curves shows the effectiveness of this efficient expression robust rejector.

A technique for computing the field of short-term glacier surface motion on a local scale is presented. Time-lapsed image negatives, digitized to a high resolution, provide the raw data for the three-stage technique. First, cross-correlation is used to establish a number of candidate displacement vectors for a series of regularly spaced templates. A relaxation-labeling routine is then applied to select the most appropriate candidate vectors, according to the local flow. Novel aspects of the relaxation algorithm include a new, efficient form of the support function and the absence of a null-match category. A new development is the application of a post filter to the resultant flow field, providing suitable displacement vectors for templates that were originally unmatched and correcting vectors that are still inconsistent with the local flow. Results from an image sequence from New Zealand's Mount Cook National Park show the superiority of the technique over the maximum cross-correlation method and demonstrate the effectiveness of the post filter in improving correlation-relaxation labeling.

Template matching is used in many practical applications for tracking the movement of clouds. To accommodate their non-rigid nature, additional constraints are often applied to produce cloud motion vectors that are more indicative of the underlying flow. One such approach is correlation-relaxation labelling, which uses a relaxation algorithm to refine sets of velocity vectors produced by template matching. The quality of the final result depends on the initial motion vectors and hence the choice of similarity metric at the template matching stage. Ordinal measures (OM) are a new class of matching functions that are sensitive to image noise. The use of OM for cloud tracking is evaluated both independently and as the first stage of a correlation-relaxation labelling scheme. Experimental results for noise-corrupted sequences show that the direct application of OM does not produce meaningful results. However, within a correlation-relaxation framework, OM produce a more consistent motion field than the widely ...

The development of 3D face recognition algorithms that are robust to variations in expression has been a challenge for researchers over the past decade. One approach to this problem is to utilize the most stable parts on the face surface. The nasal region's relatively constant structure over various expressions makes it attractive for robust recognition and, in this paper, the use of features from the three-dimensional shape of nose is evaluated. After denoising, face cropping and alignment, the nose region is cropped and 16 landmarks robustly detected on its surface. Pairs of landmarks are connected, which results in 75 curves on the nasal surface; these curves form the feature set. The most stable curves over different expressions and occlusions (due to glasses) are selected and used for nose recognition. The Bosphorus dataset is used for feature selection and FRGC v2.0 for nose recognition. Results show a rank-one recognition rate of 82.58% using only two training samples with varying expression for 505 different subjects and 4879 samples, and 90.87% and 81.61% using the FRGC v2.0 dataset Spring2003 folder for training and the Fall2003 and Spring2004 folders for testing, for neutral and varying expressions, respectively. Given the relatively low complexity of the features and matching, the varying expression results are encouraging.

Although 3D face imaging is increasingly popular, many 3D facial imaging systems have significant noise components which needs to be reduced by post-processing if meaningful recognition results are desired. For best results, the denoising algorithm must be chosen appropriately, using the noise distribution, and its parameters tuned. In this paper, a thorough analysis on the influence of different denoising techniques on the performance of various holistic 3D face recognition techniques is provided. After resampling and denoising the input data, the face is cropped, aligned and normalised. This results in the feature space which is split into two sets: gallery and query. Recognition ranks are then computed on the feature space for different denoising algorithms. Result show that, despite its simplicity, the median filter produces the best results. However, its best results are achieved using a mask size much larger than is commonly used in 3D facial denoising. For all techniques, results show that denoising can be applied using parameters significantly higher than those traditionally used as this improves the recognition performance.

Describes a new fully motion-adaptive spatio-temporal filtering technique to reduce the speckle in ultrasound images. The advantages of this approach are demonstrated in echocardiographic boundary detection and in comparison with other techniques. The first stage of many automated echocardiographic image interpretation schemes is filtering to reduce the amount of speckle noise. The authors show how the two-dimensional least mean squares (TDLMS) filter can be configured as a motion-compensated filter for a time sequence of ultrasound images that eliminates the blurring associated with direct averaging. For an image corrupted by multiplicative speckle noise, the mode of the intensity distribution approximates the maximum likelihood estimator. In consequence, the temporal filter's output is biased towards the mode from the mean, using information contained within the speckle itself. A new adaptive algorithm for controlling the filter's convergence is also included. To evaluate performance, application to simulated, phantom, and an in vivo test sequence of the carotid artery are considered in comparison with other techniques. The effect of filtering on edges is of great importance, as these are used by subsequent image interpretation schemes. Quantitative measurements demonstrate the effectiveness of the Biased TDLMS filter, for both noise reduction and edge preservation. Echocardiographic images have a high noise content and suffer from poor contrast. Despite this challenging environment, the Biased TDLMS filter is shown to produce images that are better inputs for subsequent feature extraction. The benefits for echocardiographic images are highlighted by considering the problems of mitral valve analysis and extraction of the left atrium boundary.

This paper describes a symbolic approach to relational matching. The novelty of the method lies in its Bayesian modelling of relational consistency which leads to a global matching criterion with a unique mathematical structure and robustness to error. Unlike many alternatives in the literature, the method is not limited to the use of binary constraints; it can accommodate N -ary relations of varying order. In consequence of this assumed model, the consistency of match is gauged by a compound exponential function of a higher-order Hamming distance between symbolic relations; there is a single exponential associated with each potential relational mapping. These exponential functions naturally soften the symbolic constraints represented by the relational mappings. This compound exponential structure also bestows a number of tangible benefits over the use of quadratic alternatives. In the first instance, it both renders the method more robust to errors and allows it to operate effectively in a large space of relational mappings. Moreover, this robustness to inconsistency means that the method may be operated without the need for an explicit null matching process. Unmatchable entities are identified by a constraint filtering operation once the relaxation scheme has converged. The utility of the method is illustrated on the matching of hedge structures in SAR images against their cartographic representation in a digital map.

One of the challenging issues for 3D face recognition is face alignment. Many alignment algorithms are computationally expensive, making them unsuitable for real-time biometrics, or not robust enough to detect large variations in pose. In this work, a novel algorithm for 3D face rotational alignment is proposed, that uses the nose region. After preprocessing and nose region identification, alignment is performed by applying two energy functions to the nose footprint, identified as the largest filled region in the inverted depth map. These functions are minimised using Simulated Annealing and the Levenberg-Marqurdt algorithm. The energy minimisation and segmentation procedures continue iteratively until a stopping criterion is met. The method has been applied to images from the Face Recognition Grand Challenge (FRGC) v2 dataset and the consistency of its alignment has been verified using the iterative closest point (ICP) algorithm. As a self-dependent algorithm, it does not require a pre-aligned image as a reference and also has a high computational speed, approximately three times faster than the brute force ICP technique. (6 pages)

The use of morphological scale-spaces has become a popular tool for the multiscale analysis of grey-scale images. However, in common with other morphological techniques, their extension to color and other multichannel images is not straightforward because of the absence of an unambiguous ordering. This chapter describes an approach to the development of color morphological scale-spaces using area openings and closings based on the identification and processing of vector extrema. The properties, development and implementation of the resulting color area morphology scale-spaces are described and studies of their application to color image segmentation and noise reduction are presented.

Noses are hard to conceal and relatively invariant to facial expression. Notwithstanding, their use as a biometric has been largely unexplored. Using photometric stereo images, this paper proposes two new features for nose recognition. The first of these uses Fourier descriptors to capture the ridge shape, from the nasion to the tip, and the second uses geometric ratios. Both features are robustly detected using the curvature of the surface normals to locate landmarks. Recognition results for a database of 40 individuals show that, individually, the new features out-perform an eigenface approach for an image of the nasal region. When combined they have a very respectable recognition rate for methods based on one dimensional features, indicating their potential for use within multi-feature recognition systems. (6 pages)

Connected operators are an important tool for the analysis of greyscale images. In extending them to colour and other vector images there are a number of issues that must be addressed, including the definition of extrema, region merging criterion and the preservation of idempotence. This paper reviews the recently proposed approaches to these problems and considers some of the choices which must be made in the design of effective colour connective operators. An evaluation of the noise reduction performance resulting from these choices is presented. The use of the colour connected sieves in conjunction with the watershed transform for image segmentation is also investigated.

The Bath University Matching Pursuit (BUMP) project aims at developing new matching pursuit (MP) algorithms for still image and video compression. Compared to traditional MP coding, BUMP has many new features that have been designed with the aim of producing an efficient MP representation with reduced complexity. In this paper, we study in depth the characteristics the of bit rate change along the coding process of BUMP coder, and then propose a mathematical framework for describing the relationship between the number of coded atoms and the resultant bit rate. The framework has a very simple form and demonstrates very high modeling accuracy, which allows a precise estimation of the number of bits to be used throughout the coding process. We then present an adaptive algorithm for predicting the number of bits for use in our hybrid wavelet/MP video coder. Experiments show that the proposed rate estimation scheme is able to provide an extremely accurate estimation for the BUMP based video coder.

Improved coding of atoms in image compression by matching pursuits (MP) after a wavelet decomposition is achieved. The positions of atoms in the wavelet sub-bands are communicated by run length coding. The MERGE coding scheme replaces individual coding of any alphabet of symbols by dividing the symbols into groups and run length coding their positions. It is shown theoretically that the use of MERGE with an efficient run length coder approaches the theoretical entropy of an alphabet of symbols. The embedded nature of MERGE makes it a near-ideal coding scheme for the amplitudes of atoms. An additional layer of MERGE coding for the basis indices of atoms approaches the theoretical optimum and is superior to either separable or combined Huffman coding.

A morphological gradient approach to color edge detection Evans, A.N.; Liu, X.U.; Image Processing, IEEE Transactions on Volume 15, Issue 6, June 2006 Page(s):1454 - 1463 Abstract: A new color edge detector based on vector differences is proposed. The basic technique gives as its output the maximum distance between the vectors within a mask. When applied to scalar-valued images, the method reduces to the classic morphological gradient. The technique is relatively computationally efficient and can also be readily applied to other vector-valued images. To improve the performance in the presence of noise, a novel pairwise outlier rejection scheme is employed. A quantitative evaluation using Pratt's figure of merit shows the new technique to outperform other recently proposed color edge detectors. In addition, application to real images demonstrates the approach to be highly effective despite its low complexity.

A new vector ordering is proposed and its application to colour image processing demonstrated. Partial ordering is first used rank the vectors in groups of two and then a simple reduced ordering is applied to each pair to give a full ordering. As the initial ordering is based on pairs of vectors, the technique is termed pairwise vector ordering. The new ordering is used to develop a median filter and colour morphological operations. A quantitative evaluation shows the pairwise vector median to be less effective for noise reduction than the well known vector median filter, albeit with reduced complexity. Colour morphological erosions and dilations based on the pairwise ordering are shown to produce well behaved results that conform with expectations, without any of the problems associated with applying the operations to the individual channels. Their application to multiscale image analysis is also investigated.

Morphological sieves are a popular tool for scale-space image analysis. Recent work has considered the development of colour and multichannel sieves, and their application to image segmentation. This paper proposes a new colour sieve based on the geometry of the local convex hull, providing a more flexible approach to extrema definition. Results show the new approach to have a similar segmentation performance to existing colour sieves but with an improved performance in terms of noise reduction.

Texture segmentation based on local morphological pattern spectra provides an attractive alternative to linear scale spaces as the latter suffer from blurring and do not preserve the shape of image features. However, for successful segmentation, pattern spectra derived using a number of structuring elements, often at different orientations, are required. This paper addresses this problem by using area morphology to generate a single pattern spectrum, consisting of a local granulometry and anti-granulometry, at each pixel position. As only one spectrum is produced, segmentation is performed by directly using the spectrum as the feature vector instead of taking pattern spectrum moments. Segmentation results for a simulated image of Brodatz textures and test images from the Outex texture database show the potential of the new approach.

A new approach to speckle reduction with texture preservation is presented termed granulometric shrinkage. The technique is analogous to wavelet shrinkage but with the wavelet transform replaced by a nonlinear multiscale granulometric sieve. The sieve employs area morphology, thus removing any shape bias associated with the use of fixed structuring elements. As the granulometric sieve is by nature edge preserving, the benefits of the new approach include improved edge and texture preservation. Experimental results using a SPOT image corrupted by synthetic speckle show the new approach to have a better speckle reduction capability than conventional wavelet shrinkage, with improved preservation of edges and other high frequency image features.

Morphological scale-spaces have become an important tool for analysing greyscale images. However, their extension to colour images has proven elusive until recently. In this paper an original evaluation of two recently proposed colour sieves is presented, both algorithmically and in terms of their computational and segmentation performance. A new colour sieve structure is also proposed, motivated by the relative advantages of the two sieves previously studied. A quantitative evaluation of the segmentation performance using a set of images with human ground truth from the Berkeley dataset shows the new method to produce the best segmentation performance.

Methods of reducing the complexity of the matching pursuits algorithm with minimal loss of fidelity when coding displaced frame difference (DFD) images in video compression are investigated. A full search using 2D basis functions is used as a benchmark. The use of separable 1D bases greatly reduces the complexity, and significant further reductions are achieved by using only a 1D inner product search to locate the atom position, followed by a further 1D inner product search in the opposite direction to identify the second 1D basis function. To avoid ignoring significant structures orthogonal to the search direction, it is proposed to alternate the initial search direction between horizontal and vertical scanning. This produces a modest increase in distortion compared to the full 2D search, with a complexity reduction in excess of an order of magnitude.

In video compression image noise degrades the overall compression performance. Reducing the noise content can produce a more compressible sequence, ideally with minimal effect on the visual quality. This paper presents a new approach to noise reduction using sliding window spatio-temporal attribute morphology filters with both area and power attributes. A combined spatial and spatio-temporal algorithm is proposed that produces a better overall peak signal to noise ratio (PSNR) performance than either spatial or spatio-temporal filtering individually. The compression benefits of such an approach are also demonstrated.

A new morphological gradient operator for colour images is introduced that can be viewed as a direct extension of the well known morphological gradient. In this approach, each pixel is considered as multivariate data and its output is the maximum distance between any two points within a structuring element, determined by a norm. In contrast with other nonlinear schemes, this approach reduces to the morphological gradient for single channel images. It is also computationally efficient and responds well to step edges. To overcome, any sensitivity to noise a robust colour morphological gradient operator is proposed that rejects outlying vector pairs before determining the maximum distance. Results show the effectiveness of the techniques.

A novel technique for motion description encoding is proposed. The method is based on a vector list and a mapping, each of which is losslessly encoded using variety of algorithms. Results show that in comparison to H.263, the best performing algorithms offer a significant reduction in the number of motion bits required, for motion fields with relatively few unique vectors. The bits saving are particularly important for low bit rate coding applications.

A multi-scale texture segmentation algorithm for SAR images based on the discrete wavelet transform is presented. Responses from different sub-bands are used to form a feature vector for each pixel position that is the input to the classification scheme. To further improve the classification results, the tree-structured wavelet packet transform is used to automatically identify a suitable sub-set of elements from the feature vector. Results show that this approach is effective for both the redundant and non-redundant wavelet transforms.

Ordinal measures, based on the relative order of intensity values within window, provide a robust method for finding correspondences between non-ideal images. However, for medium sized windows their discriminatory power is low. To overcome this problem, they can be applied within a correlation-relaxation labelling framework. The advantages of this approach are demonstrated in application to cloud tracking in a sequence of Meteosat infrared images.

This paper presents a computationally efficient motion estimation technique based on image sampling which determines the dominant motion between pairs of images. The technique is suited to low complexity, low bit rate multimedia applications, where the objective is to achieve good fidelity without the overhead of full motion compensation. This can be achieved if the dominant motion is a combination of translation, rotation and zoom, which can be described by a similarity transformation. The method adopts a new approach to determining the model parameters, based on generating a list of parameter estimates from pairs of block motion vectors and selecting the mean of those estimates close to the median, The method gives a good sub-pixel dominant motion estimate by sampling as little as 1/20/sup th/ of the image area. Results show the method to be accurate and robust, with low computational requirements.

Digital video pre-processing aims to eliminate the image noise that reduces the overall compression performance while preserving the visual quality of the image sequence. This paper extends our work on attribute morphology pre-processing to multidimensional spatio-temporal filters using both the standard volumetric approach and a sliding window implementation. A novel combination of spatial and spatio-temporal filters, each with their own attribute limits, is introduced and produces the best overall peak signal to noise ratio performance. The recently proposed power attribute is also shown to out-perform the area attribute.

The aim is to present a model capable of capturing the statistical correlations introduced by multiplicative noise effects in SAR image data. The motivation behind the study is the need to model noise statistics so that a Bayesian relaxation scheme may be applied to the detection of features in SAR images. The authors present a model which predicts the edge or line gradient distributions to be a product of Rayleigh and Bessel function components; this factorisation separates the correlated and uncorrelated components of the noise. >

The paper presents a method for controlling the bandwidth given to the motion description in motion-compensated video coding by restricting the number of motion vectors used. A new algorithm, called the metric method, for selecting the vectors is presented. This approach produces predicted images that are a close approximation of the optimal that can be achieved for a fixed number of vectors with relatively little additional computational cost. Results show that, in a video codec, the metric method produces lower overall errors than the standard approach of sending one vector per block, and is better than the histogram method previously used to select the vectors.

A technique for determining the full field motion of large, non-rigid bodies is presented. The technique has two stages. Firstly, a number of candidate displacement vectors are generated for a series of regularly spaced templates, using a region-based matching procedure. These multiple candidates are then smoothed using probabilistic relaxation labelling. The relaxation routine gauges the compatibility between pairs of labels for templates within a local neighbourhood. The size of the neighbourhood is determined by the Hamming distance between template centres. This approach copes well with errors in the initial matching caused by uncertainties in the correlation surface. Results are presented in application to a time sequence of high resolution digital images, obtained from photographic negatives of a glacial scene in New Zealand's Mount Cook National Park.

The authors demonstrate that relaxation labelling methods may be rendered effective in the matching of SAR data, provided that more flexible relational models are adopted. The critical ingredient of their method is the use of a hierarchical graph representation of the scene-data and the model. Here the constraints are graded according to their relational power; the more expressive and potentially more fragile perceptual groupings provide the most compatibility constraints while those with the more ambiguous, yet easily recovered, adjacency relations are much weaker. The application vehicle for the study reported in this paper is the matching of hedges detected in SAR images against their cartographic representation in a digital map. The prerequisites for successful matching are the reliable segmentation of the hedge structures and the recovery of a stable relational description suitable for matching against the map data. The former is achieved via the sequential application of a robust intensity ridge detection algorithm and contour grouping prior to the identification of meaningful linear segments. Since linear hedge structures are largely associated with the quadrilateral boundaries of fields, it is parallel and perpendicular groupings of the segments that potentially provide the most perceptual constraints for use in matching. Unfortunately, under conditions of severe noise and clutter, the reliable identification of such groupings is invariably frustrated.

A new gradient operator for greyscale images is proposed, the output of which is based on the median centred difference of the intensities within a local mask. The operator is a simplified form of the robust colour morphological gradient and, as such, is robust to impulsive noise and provides well-localised gradient estimates. A quantitative evaluation using simulated images shows that the new operator is particularly suitable for the watershed transform, producing significantly better watershed segmentation results than conventional edge operators such as the morphological gradient and the Sobel.

This paper extends speckle filtering from two to three dimensions to exploit the temporal nature of speckle to improve its reduction. A motion adaptive two-dimensional least mean square (TDLMS) filter has been applied to a time series of ultrasound images and the suitability of the results for a further computer vision stage evaluated. This filter can improve images better than by direct averaging, with a major advantage that it preserves edge data and hence fine detail in dynamic images. In order to make its operation yet more suited to speckle reduction, a novel modification to the TDLMS filter is introduced that includes a median filter within its structure. Quantitative measures are used to determine the performance of the filters on speckle reduction and in regions containing edges. Results show that the TDLMS filter produces better results than direct averaging and the modified TDLMS filter improves these further, indicating the potential for further modification.

A new nonlinear technique for filtering motion fields and other multivariate data is introduced. The method is developed from mathematical morphological area openings and uses a vector to scalar transform, in which each vector is replaced by the sum of the distances to its connected neighbours, to control the growth of extrema regions. As the filter either perfectly preserves or completely removes image components, it is able to remove noise without altering significant features. In addition, at larger area sizes a meaningful interpretation of the underlying structure is achieved. Results show that the vector area morphology sieve performs well in comparison to the widely used vector median filter.

In video compression, image pre-processing is used to improve the overall compression performance by removing noise from the image sequence. The compressibility can further be improved by removing other visually redundant information that the human visual system is not sensitive to, provided that the visual quality of the image sequence is preserved. A new approach to pre-processing is presented, based on attribute morphology in which extrema regions are removed provided that they meet some perceptual criterion, given by the attribute limit. Both the standard area and a new attribute, based on the power within the removed component, are investigated. Psychovisual experiments are used to determine the psychovisually lossless attribute limits and the limits expressed in a generalised form. The performance gain achieved by this approach is determined by comparing the codec outputs for original and the pre-processed images. Results show the compressibility of the pre-processed images, assessed by a number of compression methods, is significantly improved, thus demonstrating the advantages of the attribute morphology methods, with the power attribute providing the greatest improvement.

The authors investigate the use of filtering techniques to reduce speckle in ultrasound images, to improve their suitability for later feature extraction. The maximum likelihood estimator for a speckle corrupted image is shown to correspond to the statistical mode but this is difficult to determine for small populations, such as those contained by a filter mask. The truncated median filter approximates the mode by using the order of known image statistics and provides a fully automated image processing technique for speckle filtering. The filter's performance is established using a new quantitative evaluation scheme that closely considers the effect of filtering on edges, a key factor when applying features extraction in automated image interpretation. Application to in vivo and phantom test images shows that the truncated median filter provides clear images with strong edges, of quality exceeding that of other techniques. These benefits are confirmed by the application of feature extraction in arterial wall labelling