This paper describes a new hybrid detection method that combines exemplar based approach with discriminative patch selection. More specifically, we applied a modified random forest for retrieval of input similar local patches of stored exemplars while rejecting background patches. A recursive algorithm based on dynamic programming 2D matching optimization is applied after the aforementioned patch retrieving stage in order to enforce geometric constraints of object patches. Our proposed approach demonstrates experimentally that it performs well while maintaining the capability for incremental learning.

We address the problem of Multiple View Face Detection (MVFD) in unconstrained environments. In order to achieve generalized face detection we use part-based image representations by tessellation of small image patches, which are typified by 2D vector arrays. Faces are detected by a method named Vector Array Recognition by Indexing and Sequencing (VARIS). VARIS is designed to find the optimal similarity matching between the input image and stored exemplars while allowing wide geometrical variations that are limited only by topological constraints. Aggregated similarity is further enhanced by matching the input images with compound exemplars. The novel compounding procedure also reduces the number of exemplars necessary for each class representation. VARIS with compounding performs efficient parallel classification and has polynomial computational complexity.

In this work, we develop an efficient technique to transform a multiclass recognition problem into a minimal binary classification problem using the Minimal Classification Method (MCM). The MCM requires only log 2 N classifications whereas the other methods require much more. For the classification, we use Support Vector Machine (SVM) based binary classifiers since they have superior generalization performance. Unlike the prevalent one-versus-one strategy (the bottom-up one-versus-one strategy is called tournament method) that separates only two classes at each classification, the binary classifiers in our method have to separate two groups of multiple classes. As a result, the probability of generalization error increases. This problem is alleviated by utilizing error correcting codes, which results only in a marginal increase in the required number of classifications. However, in comparison to the tournament method, our method requires only 50% of the classifications and still similar performance can be attained. The proposed solution is tested with the Columbia Object Image Library (COIL). We also test the performance under conditions of noise and occlusion.

This paper presents a new approach for shape description and invariant recognition by geometric-normalization implemented by neural networks. The neural system consists of a shape description network, a normalization network and a recognition stage based on fuzzy pyramidal neural networks. The description network uses a novel approach for hierarchical shape segmentation and representation which expands the image shapes into localized feature tokens. These feature tokens form a compact description of the shape and its components that include information on their location, size and orientation. The description network, which is composed of a novel pyramidal architecture called the Vectorial Gradual Lattice Pyramid, processes in parallel a new vectorial scale space representation of the shape. A novel measure called Cancellation Energy is used to determine the feature tokens. The normalization network utilizes the location, size and orientation information in the feature tokens to geometric-normalize the shape or its components with respect to these parameters. The recognition network which has a pyramidal structure, uses a fuzzy representation of these normalized feature tokens to achieve robust invariant recognition. Experimental results demonstrate robust recognition in large variations of scale, rotation, translation and also in moderate affine transformations and partial occlusion.

In this paper, we develop a novel method for view-based recognition of human action/activity from videos. By observing just a few frames, we can identify the activity that takes place in a video sequence. The basic idea of our method is that activities can be positively identified from a sparsely sampled sequence of a few body poses acquired from videos. In our approach, an activity is represented by a set of pose and velocity vectors for the major body parts (hands, legs, and torso) and stored in a set of multidimensional hash tables. We develop a theoretical foundation that shows that robust recognition of a sequence of body pose vectors can be achieved by a method of indexing and sequencing and it requires only a few pose vectors (i.e., sampled body poses in video frames). We find that the probability of false alarm drops exponentially with the increased number of sampled body poses. So, matching only a few body poses guarantees high probability for correct recognition. Our approach is parallel, i.e., all possible model activities are examined at one indexing operation. In addition, our method is robust to partial occlusion since each body part is indexed separately. We use a sequence-based voting approach to recognize the activity invariant to the activity speed.

A novel method for view-based recognition of human activity is presented. The basic idea of our method is that activities can be positively identified from a sparsely sampled sequence of few body poses acquired from videos. In our approach, an activity is represented by a set of pose and velocity vectors for the major body parts (hands, legs and torso) and stored in a set of multidimensional hash tables. We show that robust recognition of a sequence of body pose vectors can be achieved by a method of indexing and sequencing and it requires only few vectors (i.e. sampled body poses in video frames). We find that the probability of false alarm drops exponentially with the increased number of sampled body poses. We also achieve speed invariant recognition by eliminating the time factor and replacing it with sequence information. Experiments performed with videos having 8 different activities show robust recognition even for different viewing directions.

This paper presents a novel hierarchical shape description scheme based on propagating the image gradient radially. This radial propagation is equivalent to a vectorial convolution with sector elements. The propagated gradient field collides at centers of convex/concave shape components, which can be detected as points of high directional disparity. A novel vectorial disparity measure called Cancellation Energy is used to measure this collision of the gradient field, and local maxima of this measure yield feature tokens. These feature tokens form a compact description of shapes and their components and indicate their central locations and sizes. In addition, a Gradient Signature is formed by the gradient field that collides at each center, which is itself a robust and size-independent description of the corresponding shape component. Experimental results demonstrate that the shape description is robust to distortion, noise and clutter. An important advantage of this scheme is that the feature tokens are obtained pre-attentively, without prior understanding of the image. The hierarchical description is also successfully used for similarity-invariant recognition of 2D shapes with a multi-dimensional indexing scheme based on the Gradient Signature.

In this paper, we develop a novel framework for robust recovery of three-dimensional (3-D) surfaces of faces from single images. The underlying principle is shape from recognition, i.e., the idea that pre-recognizing face parts can constrain the space of possible solutions to the image irradiance equation, thus allowing robust recovery of the 3-D structure of a specific part. Parts of faces like nose, lips and eyes are recognized and localized using robust expansion matching filter templates under varying pose and illumination. Specialized backpropagation based neural networks are then employed to recover the 3-D shape of particular face parts. Representation using principal components allows to efficiently encode classes of objects such as nose, lips, etc. The specialized networks are designed and trained to map the principal component coefficients of the part images to another set of principal component coefficients that represent the corresponding 3-D surface shapes. To achieve robustness to viewing conditions, the network is trained with a wide range of illumination and viewing directions. A method for merging recovered 3-D surface regions by minimizing the sum squared error in overlapping areas is also derived. Quantitative analysis of the reconstruction of the surface parts in varying illumination and pose show relatively small errors, indicating that the method is robust and accurate. Several examples showing recovery of the complete face also illustrate the efficacy of the approach.

This paper presents a novel approach for detection and segmentation of man made generic shapes in cluttered images. The set of shapes to be detected are members of affine transformed versions of basic geometric shapes such as rectangles, circles etc. The shape set is represented by its vectorial edge map transformed over a wide range of affine parameters. We use vectorial boundary instead of regular boundary to improve the robustness to noise, background clutter and partial occlusion. Our approach consists of a detection stage and a verification stage. In the detection stage, we first derive the energy from the principal eigenvectors of the set. Next, an a posteriori probability map of energy distribution is computed from the projection of the edge map representation in a vectorial eigen-space. Local peaks of the posterior probability map are located and indicate candidate detections. We use energy/probability based detection since we find that the underlying distribution is not Gaussian and resembles a hypertoroid. In the verification stage, each candidate is verified using a fast search algorithm based on a novel representation in angle space and the corresponding pose information of the detected shape is obtained. The angular representation used in the verification stage yields better results than a Euclidean distance representation. Experiments are performed in various interfering distortions, and robust detection and segmentation are achieved.

A novel framework for the recovery of 3D surfaces of faces from single images is developed in this work. The underlying principle is shape from recognition, i.e. the idea that pre-recognizing face parts can constrain the space of possible solutions to the image irradiance equation, thus allowing robust recovery of the 3D structure of a specific part. Generic face parts are localized using expansion matching filters. Specialized backpropagation based neural networks are then employed in the recovery of the recognized and localized face part. This circumvents directly solving the image irradiance equation. Instead, the relationships between variations in face structure and appearance under varying pose and illumination is learned. Representation using principal components allows to efficiently encode classes of objects such as nose, lips, etc. for association with the neural networks. Quantitative analysis of the reconstruction of the surface parts show relatively small errors, indicating that this system can accurately recover 3D face structure from single images invariant to pose and illumination.

This paper presents a novel approach for detection and segmentation of generic shapes in cluttered images. We use vectorial eigenvectors to compactly represent a large set of possible appearances of primitive shapes. A posterior energy probability map of the image is calculated in the vectorial eigenspace to yield a relative similarity measure. The detection of genetic shapes is realized by detecting local peaks of the probability map. We find that eigenspace energy is more suitable for representation of sparse sets such as our affine set. At each local probability maxima, a fast search approach based on a novel representation by an angle space is employed to determine the best matching between models and the underlying sub-image. We find that angular representation in multidimensional search corresponds better to Euclidean distance than conventional projection and yields improved classification of noisy shapes. Experiments are performed in various interfering distortions, and robust detection and segmentation are achieved.

This correspondence describes a novel approach to three-dimensional (3-D) motion estimation of planar objects based on eigen-normalization, expansion matching (EXM), and a scaled orthographic projection model. Our approach leads to a comprehensive temporal description of all the degrees of freedom in 3-D (three rotations and three translations). Experiments with video streams show robust estimation of the real 3-D rotations and translations of the objects in motion.

In this paper a novel framework for the recovery of 3D surfaces of faces from single images is developed. The underlying principle is shape from recognition, i.e. the idea that pre-recognizing face parts can constrain the space of possible solutions to the image irradiance equation, thus allowing robust recovery of the 3D structure of a specific part. Shape recovery of the recognized part is based on specialized backpropagation based neural networks, each of which is employed in the recovery of a particular face part. Representation using principal components allows to efficiently encode classes of objects such as nose, lips, etc. The specialized networks are designed and trained to map the principal component coefficients of the shading images to another set of principal component coefficients that represent the corresponding 3D surface shapes. A method for integrating recovered 3D surface regions by minimizing the sum squared error in overlapping areas is also derived. Quantitative analysis of the reconstruction of the surface parts show relatively small errors indicating that the method is robust and accurate. The recovery of a complete face is performed by minimal squared error merging efface parts.

This paper presents a novel approach for detection and segmentation of generic shapes in cluttered images. The underlying assumption is that generic objects that are man made, frequently have surfaces which closely resemble standard model shapes such as rectangles, semi-circles etc. Due to the perspective transformations of optical imaging systems, a model shape may appear differently in the image with various orientations and aspect ratios. The set of possible appearances can be represented compactly by a few vectorial eigenbases that are derived from a small set of model shapes which are affine transformed in a wide parameter range. Instead of regular boundary of standard models, we apply a vectorial boundary which improves robustness to noise, background clutter and partial occlusion. The detection of generic shapes is realized by detecting local peaks of a similarity measure between the image edge map and an eigenspace combined set of the appearances. At each local maxima, a fast search approach based on a novel representation by an angle space is employed to determine the best matching between models and the underlying subimage. We find that angular representation in multidimensional search corresponds better to Euclidean distance than conventional projection and yields improved classification of noisy shapes. Experiments are performed in various interfering distortions, and robust detection and segmentation are achieved.

A novel generalized feature extraction method based on the expansion matching (EXM) method and on the Karhunen-Loeve transform (KLT) is presented. The method provides an efficient way to locate complex features of interest like corners and junctions with reduced number of filtering operations. The EXM method is used to design optimal detectors for a set of model elementary features. The KL representation of these model EXM detectors is used to filter the image and detect candidate interest points from the energy peaks of the eigen coefficients. The KL coefficients at these candidate points are then used to efficiently reconstruct the response and differentiate real junctions and corners from arbitrary features in the image. The method is robust to additive noise and is able to successfully extract, classify, and find the myriad compositions of corner and junction features formed by combinations of two or more edges or lines. This method differs from previous works in several aspects. First, it treats the features not as distinct entities, but as combinations of elementary features. Second, it employs an optimal set of elementary feature detectors based on the EM approach. Third, the method incorporates a significant reduction in computational complexity by representing a large set of EXM filters by a relatively small number of eigen filters derived by the KL transform of the basic EXM filter set. This is a novel application of the KL transform, which is usually employed to represent signals and not impulse responses as in our present work.

This paper presents a novel method, based on magnetic field principles, for obtaining the axes of shapes. The method is based on directional information of the shape's boundary. By simulating a parallel algorithm, we are able to generate the inner as well as the outer axes (axes of concavities) of the shape. The preprocessing phase for this algorithm involves obtaining the shape's gradient. Each point of the gradient is substituted by a minute magnetic dipole. The cumulative magnetic field due to these dipoles is accumulated at all points in the image in a one-pass algorithm. The magnitude of the final magnetic vector field has valleys that are created from mutual and directionally balanced cancellations of opposing boundary segments. These valleys signify the axes of the shape. The axes are obtained by performing a valley search. The magnetic field modeling (MFM) method has an advantage over previous approaches since it utilizes not only the location information of the boundary, but also its directional information. As demonstrated, experimental results of the MFM method are much improved, compared to other skeletonization algorithms which tend to generate spurious and noisy axes.

A framework for the reconstruction of smooth surface shapes from shading images is presented. The method is based on using a backpropagation based neural network for learning brightness patterns and associating them with range data. The network is designed to reconstruct surface range from localized intensity patches of 7/spl times/7 pixels. Two methods for training the network are investigated, one based on a novel weight diffusion process which enforces a local smoothness constraint and the other using the eigen coefficients of the input and output patterns which make the training computationally efficient. An elegant and simple method for integrating reconstructed surface patches by minimizing the sum squared error in overlapped areas is derived. Results are shown for reconstruction of simple shapes like cylinders, hyperboloids and paraboloids as well as complex shapes like facial structure from intensity images.

A novel method for extracting parametric junction and corner features in images is presented. By treating each complex feature as a combination of elementary line and edge features, the method provides an efficient way to locate features of interest with reduced number of filtering operations. The expansion matching (EXM) method is used to design optimal detectors for a set of elementary model shape features. Next, the principal components of the Karhunen-Loeve (KL) representation of these model EXM detectors are used to filter the image and extract candidate interest points derived from the energy peaks of the eigen coefficients. The KL coefficients at these candidate points are then used to efficiently reconstruct the response and differentiate real junctions and corners from arbitrary features in the image. The method is robust to additive noise and is able to successfully extract, classify and find the myriad compositions of corner and junction features formed by combinations of two or more elementary edges or lines.

This paper describes a novel approach to 3D motion estimation of planar objects based on eigen-normalization, expansion matching (EXM) and a scaled orthographic projection model. Our approach leads to a comprehensive temporal description of all degrees of freedom in 3D (3 rotations and 3 translations). The 3D motion parameters of the objects are approximated by the corresponding affine parameters. The objects in each frame of a video sequence are normalized to a set of canonical images using principal component normalization procedure. The normalization approach here is based on principal components of the intensity weighted spatial values and not on the intensity values as in works such as eigenfaces. The canonical images generated differ only in orientation. Expansion matching (EXM) is then used to find the differences in orientation. Affine transformations between the shapes also are derived. The pose of the shape in 3D space can therefore be estimated. Experiments on video sequences of planar and quasi-planar objects show robust estimation of the real 3D rotations and translations of the objects in motion.

A novel method for representing 3D objects that unifies viewer and model centered object representations is presented. A unified 3D frequency-domain representation, called volumetric frequency representation (VFR), encapsulates both the spatial structure of the object and a continuum of its views in the same data structure. The frequency-domain image of an object viewed from any direction can be directly extracted employing an extension of the projection slice theorem, where each Fourier-transformed view is a planar slice of the volumetric frequency representation. The VFR is employed for pose-invariant recognition of complex objects, such as faces. The recognition and pose estimation is based on an efficient matching algorithm in a four-dimensional Fourier space. Experimental examples of pose estimation and recognition of faces in various poses are also presented.

Describes an efficient approach to pose invariant pictorial object recognition employing spectral signatures of image patches that correspond to object surfaces which are roughly planar. Based on singular value decomposition (SVD), the affine transform is decomposed into slant, tilt, swing, scale, and 2D translation. Unlike previous log-polar representations which were not invariant to slant, our log-log sampling configuration in the frequency domain yields complete affine invariance. The images are preprocessed by a novel model-based segmentation scheme that detects and segments objects that are affine-similar to members of a model set of basic geometric shapes. The segmented objects are then recognized by their signatures using multidimensional indexing in a pictorial dataset represented in the frequency domain. Experimental results with a dataset of 26 models show 100 percent recognition rates in a wide range of 3D pose parameters and imaging degradations: 0-360/spl deg/ swing and tilt, 0-82/spl deg/ of slant, more than three octaves in scale change, window-limited translation, high noise levels (0 dB), and significantly reduced resolution (1:5).

A novel method for representing 3-D objects that unifies viewer and model centered object representations is presented. A unified 3-D frequency-domain representation (called volumetric/iconic spectral signatures-V/ISS) encapsulates both the spatial structure of the object and a continuum of its views in the same data structure. The frequency-domain image of an object viewed from any direction can be directly extracted employing an extension of the projection slice theorem, where each Fourier-transformed view is a planar slice of the volumetric frequency representation. The V/ISS representation call be employed for pose-invariant recognition of complex objects such as faces. The recognition and pose estimation is based on an efficient matching algorithm in a four dimensional Fourier space. Experimental examples of pose estimation and recognition of faces are also presented.

This paper describes an efficient approach to pose invariant object recognition employing pictorial recognition of image patches. A complete affine invariance is achieved by a representation which is based on a new sampling configuration in the frequency domain. Employing Singular Value Decomposition (SVD), the affine transform is decomposed into slant, tilt, swing, scale and 2D translation. From this decomposition, we derive an affine invariant representation that allows to recognize image patches that correspond to object surfaces which are roughly planar-invariant to their pose in space. The representation is in the form of Spectral Signatures that are derived from a set of Cartesian logarithmic-logarithmic (log-log) sampling configuration in the frequency domain. Unlike previous log-polar representations which are not invariant to slant (i.e. foreshortening only in one direction), our new configuration yields complete affine invariance. The proposed log-log configuration can be employed both globally or locally by a Gabor or Fourier transforms. Local representation enables to recognize separately several objects in the same image. The actual signature recognition is performed by multidimensional indexing in a pictorial dataset represented in the frequency domain. The recognition also provides 3D pose information.

A novel method unifying viewer and model centered approaches for representing structurally complex 3-D objects like human faces is presented. The unified 3-D frequency-domain representation (called volumetric/iconic spectral signatures-V/ISS) encapsulates both the spatial structure of the object and a continuum of the projection slice theorem is used to directly extract the frequency-domain image of an object as viewed from any direction. Each such Fourier-transformed view is a planar slice of the volumetric frequency representation. The V/ISS representation is employed for pose-invariant recognition of complex objects such as faces. The recognition and pose estimation is based on an efficient matching algorithm in a four dimensional Fourier space. Experimental examples of pose estimation and recognition of faces are presented.

This paper presents an efficient scheme for affine-invariant object recognition. Affine invariance is obtained by a representation which is based on a new sampling configuration in the frequency domain. We discuss the decomposition of affine transform into slant, tilt, swing, scale and 2D translation by applying singular value decomposition (SVD). The affine invariant spectral signatures (AISS) are derived from a set of Cartesian logarithmic-logarithmic (log-log) sampling configuration in the frequency domain. The AISS enables the recognition of image patches that correspond to roughly planar object surfaces-regardless of their poses in space. Unlike previous log-polar representations which are not invariant to slant (i.e. foreshortening only in one direction), the AISS yields a complete affine invariance. The proposed log-log configuration can be employed either by a global Fourier transform or by a local Gabor transform. Local representation enables one to recognize separately several objects in the same image. The actual signature recognition is performed by multi-dimensional indexing in a pictorial dataset. 3D pose information is also derived as a by-product.

This paper presents a new approach for object recognition using affine-invariant recognition of image patches that correspond to object surfaces that are roughly planar. A novel set of affine-invariant spectral signatures (AISSs) are used to recognize each surface separately invariant to its 3D pose. These local spectral signatures are extracted by convolving the image with a novel configuration of Gaussian kernels. The spectral signature of each image patch is then matched against a set of iconic models using multi-dimensional indexing (MDI) in the frequency domain. Affine-invariance of the signatures is achieved by a new configuration of Gaussian kernels with modulation in two orthogonal axes. The proposed configuration of kernels is Cartesian with varying aspect ratios in two orthogonal directions. The kernels are organized in subsets where each subset has a distinct orientation. Each subset spans the entire frequency domain and provides invariance to slant, scale and limited translation. The complete set of orientations is utilized to achieve invariance to rotation and tilt. Hence, the proposed set of kernels achieve complete affine-invariance.

A novel generalized feature extraction method based on the expansion matching (EXM) method and the Karhunen-Loeve (KL) transform is presented. This yields an efficient method to locate a large variety of features with a single pass of parallel filtering operations. The EXM method is used to design optimal detectors for different features. The KL representation is used to define an optimal basis for representing these EXM feature detectors with minimum truncation error. Input images are then analyzed with the resulting KL bases. The KL coefficients obtained from the analysis are used to efficiently reconstruct the response due to any combination of feature detectors. The method is successfully applied to real images and extracts a variety of arc and edge features as well as more complex junction features formed by combining two or more arcs or line features.

This paper presents a new approach for object recognition using affine-invariant recognition of image patches that correspond to object surfaces that are roughly planar. A novel set of affine-invariant spectral signatures (AISSs) are used to recognize each surface separately invariant to its 3D pose. These local spectral signatures are extracted by correlating the image with a novel configuration of Gaussian kernels. The spectral signature of each image patch is then matched against a set of iconic models using multidimensional indexing (MDI) in the frequency domain. Affine-invariance of the signatures is achieved by a new configuration of Gaussian kernels with modulation in two orthogonal axes. The proposed configuration of kernels is Cartesian with varying aspect ratios in two orthogonal directions. The kernels are organized in subsets where each subset has a distinct orientation. Each subset spans the entire frequency domain and provides invariance to slant, scale and limited translation. The complete set of orientations is utilized to achieve invariance to rotation and tilt. Hence, the proposed set of kernels achieve complete affine-invariance.

A novel and efficient generalized feature extraction method is presented based on the expansion matching (EXM) method and the Karhunen-Loueve (KL) transform. The EXM method is used to design optimal detectors for different features. The KL representation is used to define an optimal basis for representing these EXM feature detectors with minimum truncation error. Input images are then analyzed with the resulting KL basis set. The KL coefficients obtained from the analysis are used to efficiently reconstruct the response due to any combination of feature detectors. The method is applied to real images and successfully extracts a variety of arc and edge features as well as complex junction features formed by combining two or more arc or line features.

In practical images, ideal step edges are actually transformed into ramp edges, due to the general low pass filtering nature of imaging systems. This paper discusses the application of the expansion matching (EXM) method for optimal ramp edge detection. EXM optimizes a novel matching criterion called discriminative signal-to-noise ratio (DSNR) and has been shown to robustly recognize templates under conditions of noise, severe occlusion, and superposition. We show that our ramp edge detector performs better than the ramp detector obtained from Canny's criteria in terms of DSNR and is relatively easier to derive for various noise levels and slopes.

A computational model of auditory localization based on processing spectral amplitude cues for localizing broadband high frequency sound sources is presented. The cues extracted are binaural spectral level di erence patterns of the Head Related Transfer Functions (HRTF) corresponding to the direction of the sound source. Four di erent pattern classi ers are used to evaluate the feasibility of localization using such spectral cues. The four classi ers investigated are normalized correlation, maximum likelihood estimation, a backpropagation based fuzzy neural network model and a novel linear ltering method which optimizes a new Discriminative Matching Measure (DMM). The DMM is de ned to quantify the relative ability of each of these classi ers to identify such patterns in the presence of noise or with missing data. An independent estimate of the localization acuity using the RMS error is also made. The results indicate that both optimal DMM ltering and the fuzzy neural network model are successful in extracting azimuthal location of broadband high frequency sound sources. The proposed model attempts to complement current auditory localization theory by exploring the role of HRTF based binaural spectral level di erence cues in localizing high frequency broadband auditory stimuli.

Humans can perceive their spatial surroundings both by the visual and the auditory senses. Even though the auditory system has much lower spatial resolution, it can still serve as an alternative or supplementary information channel for the visual modality. We investigate the feasibility of conveying elaborate visual information by auditory patterns. A system is developed that transforms 2-D binary images into "auditory images." Such images are based on slow raster scanning of a perceptual auditory surface around the subject while emitting sounds that correspond to the brightness level of the images at each location. The parameters involved in such a system are quantitatively investigated with respect to their effect on performance. These parameters include resolution, level difference, sound color contrast, speed, surface curvature, etc. The graphic percepts of synthetic auditory images that correspond to simple shapes are also analyzed. The experimental results show that sound localization can be used to convey visual information quite successfully for simple shapes and low-resolution patterns.

This paper presents a novel hierarchical shape description scheme based on propagating the gradient of the image. The propagated gradient field collides at centers of convex/concave shape components, which can be detected as points of high directional disparity. A novel vectorial disparity measure called cancelation energy is used to measure this collision of the gradient field, and local maxima of this measure yield feature tokens. These feature tokens form a compact description of shapes and their components and indicate their central location and size. In addition, a gradient signature is formed by the gradient field that collides at each center, which is itself a robust and size-independent description of the corresponding shape component. Experimental results demonstrate that the shape description is robust to distortion, noise and clutter. An important advantage of this scheme is that the feature tokens are obtained pre-attentively, without prior understanding of the image. The hierarchical description is also successfully used for similarity-invariant recognition of 2D shapes with a multi-dimensional indexing scheme based on the gradient signature.

In practical images, ideal step edges are actually transformed into exponential ramp edges, due to the general low pass filtering nature of imaging systems. This paper discusses the application of a newly developed expansion matching method for optimal ramp edge detection. Expansion matching optimizes a novel matching criterion called discriminative signal to noise ratio (DSNR). The DSNR criterion represents the desirable qualities of a sharp matching response with good localization and minimal off-center response. These requirements are consistent with the three criteria of signal-to-noise ratio, localization, and multiple response suppression used by Canny (1986) and others for optimal edge detection. We compare the optimal ramp edge detector based on DSNR with the ramp edge detector derived from Canny's criteria. We show that our ramp edge detector performs better than the ramp detector obtained from Canny's criteria in terms of DSNR and is relatively easier to derive for various amounts of noise and slopes.

A novel approach for auditory localization of broadband high frequency acoustic stimuli is presented. Computational models for extracting localization cues from estimations of the incoming sound spectra at the two ears are developed. These models can be structured as feedforward networks to classify binaural ratio patterns of head related transfer functions (HRTF). The computational models studied are pattern classifiers based on normalized correlation, fuzzy neural networks and a novel linear filtering method which optimizes a new discriminative matching measure (DMM). The DMM is defined to quantify the relative ability of each of these classifiers to identify such patterns in the presence of noise. Both optimal DMM matching and fuzzy neural networks prove successful in localization with good DMM and acceptably small RMS errors.

Discusses the application of a newly developed expansion matching method for edge detection. Expansion matching optimizes a novel matching criterion called the discriminative signal-to-noise ratio (DSNR) and has been shown to robustly recognize templates under conditions of noise, severe occlusion and superposition. The DSNR criterion is better suited to evaluate matching in practical conditions than the traditional SNR since it considers as "noise" even the off-center response of the filter to the template itself. We introduce a family of optimal DSNR edge detectors based on the expansion filter for several edge models. For step edges, the optimal DSNR step expansion filter (SEF) is compared with the widely used Canny edge detector (CED). Experimental comparisons show that our edge detector yields better performance than the CED in terms of DSNR even under very adverse noise conditions. As for boundary detection, the SEF consistently yields higher figures of merit than the CED on a synthetic binary image over a wide range of noise levels. Results also show that the design parameters of size or width of the SEF are less critical than the CED variance. This means that a single scale of the SEF spans a larger range of input noise than a single scale of the CED. Experiments on a noisy image reveal that the SEF yields less noisy edge elements and preserves structural details more accurately. On the other hand, the CED output has better suppression of multiple responses than the corresponding SEF output.>

The paper describes a multiple-template generalization of a newly developed approach for template matching by signal expansion into a set of non-orthogonal template-similar basis functions. The single-template method is proven to be equivalent to "restoration" of undegraded images using the Wiener filter and optimizes a new and more practically defined matching quality criterion that the authors call discriminative signal-to-noise ratio (DSNR). Compared to the widely used matched filtering approach (also known as correlation matching) which is based an projection, expansion matching is based on decomposition and is shown to be more robust in conditions of noise, superposition and severe occlusion. In the paper, the authors extend the DSNR optimization approach to include more than one template. The generalized expansion filter presented is optimal in terms of DSNR and can be designed to elicit any desired response for each of the templates, while optimizing the DSNR criterion. The approach considers additive noise as a parameter and leads to a general formulation, of which many previous approaches (such as the synthetic discriminant function) form special cases. In the case of a single template, the formulation reverts to the previously mentioned Wiener restoration filter.>

An approach for template matching by signal expansion into a set of nonorthogonal template-similar basis functions and its generalization to multiple templates is described. The single-template method is proven to be equivalent to restoration of undergraded images using the Wiener filter and optimizes a practically defined matching criterion called discriminative signal-to-noise ratio (DSNR). Compared to the matched filtering approach which optimizes the SNR, expansion matching is more robust in conditions of noise, superposition, and severe occlusion. The multiple-template expansion filter presented is generalized to match more than one template. It is also optimal in terms of DSNR. It can be designed to elicit any desired response for each of the templates.>

Expansion matching (EXM) optimizes a novel matching criterion called discriminative signal-to-noise ratio (DSNR) and robustly recognizes templates under conditions of noise, severe occlusion and superposition. A family of optimal DSNR edge detectors is introduced based on the expansion filter for any given edge model. Experimental comparisons show that the authors' step expansion filter (SEF) yields better results than the Canny edge detector (CED) in terms of DSNR, even under adverse noise conditions. As for segmentation quality, the SEF also yields higher figures of merit than the CED over a wide range of noise levels. Experiments on real images reveal that the SEF yields less noisy edge elements and preserves structural details accurately. EXM is also effective for feature extraction.>

A novel approach to template matching using nonorthogonal expansion with template-similar basis functions is described. The results show that this expansion is superior in comparison with the normalized-correlation (matched-filtering) approach. It is proved that the expansion can be efficiently implemented by restoration filtering and that this method yields optimal discriminative signal-to-noise ratio. It is further proved that the nonorthogonal expansion can be performed with an almost unlimited variety of basis functions. Results show that the expansion approach has definite advantages over correlation, especially in conditions of noise and severe occlusion. When recognition is required in audio, radar, or sonar signals, where the principle of occlusion is replaced by superposition, correlation techniques are even more ineffective and the expansion approach yields better results.>

Presents a novel approach for template recognition by signal expansion into a set of nonorthogonal template-similar basis functions (wavelets). It is shown that expansion matching is a special case of the general nonorthogonal expansion and is equivalent to 'restoration' of undegraded images. Matching by expansion is quite robust in conditions of noise, superposition and severe occultation. Expansion matching also maximizes a new and more practically defined discriminative signal-to-noise ratio (DSNR). It is proved that maximizing the DSNR is equivalent to minimum squared error restoration by Wiener filters. Experimental comparisons with the widely used correlation matching (matched filtering), show that expansion matching yields much higher DSNR.>

Two novel probabilistic models for viewed angles and distances are derived using an observability sphere method. The method, which is based on the assumption that the prior probability density is isotropic for all viewing orientations, can be used for the computation of observation probabilities for object's aspects, features, and probability densities of their quantitative attributes. Using the sphere, it is discovered that the probability densities of viewed angles, distances, and even projected curvature have sharp peaks at their original values. From this peaking effect, it is concluded that in most cases, the values of angles and distances are being altered only slightly by the imaging process, and they can still serve as a strong cue for model-based recognition. The probabilistic models for 3-D object recognition from monocular images are used. To form the angular elements that are needed, the objects are represented by their linear features and specific points primitives. Using the joint density model of angles and distances, the probabilities of initial matching hypotheses and mutual information coefficients are estimated. These results are then used for object recognition by optimal matching search and stochastic labeling schemes. Various synthetic and real objects are recognized by this approach.>

Two novel probabilistic models for viewed angles and distances are derived by using a probability sphere method. The method is based on the assumption that the a priori probability density is isotropic for all viewing orientations of the scene. From these models two rules are suggested that deal with imaging of angles and distances. Rule A (B): there is high probability that the ratio of the scene angles (distances) to their imaged angles (distances) is closed to unity (the unique scale factor). These rules are applied to the recognition of 3-D objects which are represented by their linear features primitives. The parameters of the stochastic labeling algorithm, that is used for the recognition are estimated from angles and distances using both models. Various synthetic and real objects have been recognized by this approach.>