Abhijit Mahalanobis

Minimum squared error synthetic discriminant functions

A new synthetic discriminant function (SDF) design approach is presented that yields the best approximation of arbitrary output correlation shapes in the minimum squared error (MSE) sense. We term such filters as MSE-SDFs. Simulation results are presented to illustrate the advantages of MSE-SDFs. Also, we show that MSE-SDFs generalize minimum average correlation energy filters.

Application of correlation filters for texture recognition

We propose a new statistical method to design spatial filters to recognize and to discriminate between various textures. Unlike existing correlation filters, the proposed filters are not meant to recognize specific shapes or objects. Rather, they discriminate between textures such as terrains, background surfaces, and random image fields. The filters do not require any on-line statistical computations for extracting texture information. Therefore optical (or digital) correlators can be used for fast real-time texture recognition without segmentation. The procedure is based on the assumption that textures can be modeled as stationary random processes over limited regions of an image. The optimum filter coefficients are determined by use of eigenvector analysis. Several examples are given to illustrate the proposed scheme.

Unified framework for the synthesis of synthetic discriminant functions with reduced noise variance and sharp correlation structure

An inherent implementational limitation of the recently developed minimum average correlation energy (MACE) filters arises from the circular correlation implicit in the discrete frequency domain approach followed. An alternative methodology that uses space domain computations to overcome these limitations is presented. An improved synthetic discriminant function (SDF) that satisfies the conflicting requirements of reduced noise variance and sharp correlation peaks to facilitate ease of detection is developed by employing a unified framework for design. A quantitative evaluation of the performance characteristics of the new filter is conducted and is shown to compare favorably with the well known nminimum variance SDF and the space domain MACE filter, which are special cases of the present design.

Recent advances in distortion-invariant correlation filter design

In recent years, several new techniques have been proposed for the synthesis and implementation of correlation filters which are philosophically quite different from earlier methods. This paper provides an updated overview of the design to correlation filters and summarizes these newer techniques and their advantages.

Distance classifier correlation filters

The performance of shift-invariant distance classifiers based on correlation filters is evaluated. First, the effect of noise on a classifier designed to recognize synthetic aperture radar (SAR) is observed. Then, a 2-class ATR designed to recognize infrared images of actual targets is evaluated. The results attest to the ability of the distance classifier to tolerate distortions, and recognize targets in the presence of noise and clutter.

Distance classifier correlation filters for distortion tolerance, discrimination, and clutter rejection

A new approach to correlation filters based on quadratic distance calculations is described. The problem of distortion tolerance is addressed in terms of similarity measures. Discrimination is simultaneously addressed by optimizing the filters to maximally separate the classes. Mathematically, filter synthesis requires the inversion of diagonal matrices in the frequency domain and is a generalization of the MACE idea. The approach is shift-invariant, does not require feature extraction or image registration, and is significantly different from traditional pattern recognition techniques such as the Fisher LDF. The proposed approach is also suitable for the rejection of unknown clutter. Since recognition is based on similarity, clutter and false images which exhibit `large distances from true target classes, are easily rejected. With improved recognition and discrimination performance, and low false alarm rates, the proposed distance classifier is a promising method for multiclass target recognition in cluttered environments.

Correlation filters for texture recognition and applications to terrain-delimitation in wide-area surveillance

Terrain-delimitation is an important component of wide-area-surveillance with applications to battlefield terrain and agricultural terrain. Recently a statistical method was proposed by Mahalanobis and Singh [1] to design spatial filters to recognize and discriminate between various textures. We extend the technique for delimitation-through-texture-discrimination in SAR images. Spatial correlation filters are used for texture distinction. The filters are implementable as optical (or digital) correlators for fast real-time texture recognition without segmentation. The filter coefficients are determined via eigenvector analysis. Examples will be given to illustrate the proposed scheme for terrain-discrimination in SAR images.<<ETX>>

Quadratic distance classifier for multiclass SAR ATR using correlation filters

The recognition of targets in synthetic aperture radar (SAR) imagery using a quadratic classifier is proposed. Correlators are used to compute distances under an optimum transform to measure similarity between ideal reference images and the actual data. The transform is a filter which responds to features specifically useful for discrimination. This is attractive for model based training since only a similarity in features is required between the actual images and their class models rather than a precise match in pixel values. The quadratic terms are unaffected by shifting of the input image while linear terms are computed using shift-invariant correlation. The system is thus non-linear but shift-invariant. Specifically, distance vectors are generated by the filter banks which are analyzed by a rudimentary rule base to determine whether the input is a target image or clutter. In this paper, we describe a SAR automatic target recognizer (ATR) with results for 3 and 5 class problems. the data used is actual SAR imagery of military targets.

Synthetic discriminant functions using relaxed constraints

Historically synthetic discriminant functions have been designed using specific constraints in the encoding methodologies. In this paper we consider several design alternatives which eliminate the constraints yet appear to perform equally well with the classical versions. These new encoding methods are also found to require a significant reduction in computation with inherently better encoding accuracy. In addition, we discuss a very simple encoding method that allows for a significant increase in the number of references that can be incorporated in a filter.

Combining multiple correlators using neural networks

Designing a pattern classifier remains a difficult problem especially in the presence of noise and other degradations. Combination of multiple classifiers appears to be a good way of retaining the strengths of different classifiers while avoiding their weaknesses. Different combination schemes were proposed in the literature. As a special case of combining multiple classifiers, we consider combining correlators. Correlators are attractive for use in Automatic Target Recognition systems. Many correlation filter designs have been developed, each with its own features. Some filter designs maximize noise tolerance but do not provide sharp peaks. On the other hand, some correlation filters yield sharp correlation peaks but are overly sensitive to input noise. In this research effort, we explore the use of artificial neural network as a tool for combining correlators. Results of this implementation show improvements and indicate that combination of multiple correlators can potentially improve the classification performance.