Rohit Patnaik

SAR classification and confuser and clutter rejection tests on MSTAR ten-class data using Minace filters

This paper presents the status of our SAR automatic target recognition (ATR) work on the Moving and Stationary Target Acquisition and Recognition (MSTAR) public database using the minimum noise and correlation energy (MINACE) distortion-invariant filter (DIF). In our previous work, we used the MSTAR public database benchmark three-class problem and demonstrated better results than all prior work. In this paper, we address classification (including variants) and object and clutter rejection tests on the more challenging MSTAR ten-class public database. The Minace algorithm is shown to generalize well to this larger classification problem. We use several filters per object, but fewer DIFs per object than prior work did. We use our autoMinace algorithm that automates selection of the Minace filter parameter c and selection of the training set images to be included in the filter. No confuser, clutter, or test set data are present in the training or the validation set. In tests, we do not assume that the test inputs pose is known (as most prior work does), since pose estimation of SAR objects has a large margin of error. We also address tests with proper use of SAR pose estimates in MSTAR recognition and the use of multilook SAR data to improve performance.

Illumination invariant face recognition and impostor rejection using different MINACE filter algorithms

A face recognition system that functions in the presence of illumination variations is presented. It is based on the minimum noise and correlation energy (MINACE) filter. A separate MINACE filter is synthesized for each person using an automated filter-synthesis algorithm that uses a training set of illumination differences of that person and a validation set of a few faces of other persons to select the MINACE filter parameter c. The MINACE filter for each person is a combination of training images of only that person; no false-class training is done. Different formulations of the MINACE filter and the use of two different correlation plane metrics: correlation peak value and peak-to-correlation plane energy ratio (PCER), are examined. Performance results for face verification and identification are presented using images from the CMU Pose, Illumination, and Expression (PIE) database. All training and test set images are registered to remove tilt bias and scale variations. To evaluate the face verification and identification systems, a set of impostor images (non-database faces) is used to obtain false alarm scores (PFA).

MSTAR object classification and confuser and clutter rejection using Minace filters

This paper presents the status of our SAR automatic target recognition (ATR) work on the Moving and Stationary Target Acquisition and Recognition (MSTAR) public database using the minimum noise and correlation energy (MINACE) distortion-invariant filter (DIF). We use a subset of the MSTAR public database for the benchmark three-class problem and we address confuser and clutter rejection. To handle the full 360° range of aspect view in MSTAR data, we use a set of Minace filters for each object; each filter should recognize the object (and its variants) in some angular range. We use fewer DIFs per object than prior work did. The Minace parameter c trades-off distortion-tolerance (recognition) versus discrimination (confuser/clutter rejection) performance. Our filter synthesis algorithm automatically selects the Minace filter parameter c and selects the training set images to be included in the filter, so that the filter can achieve both good recognition and good confuser and clutter rejection performance; this is achieved using a training and validation set. In our new filter synthesis method, no confuser, clutter, or test set data are used. The peak-to-correlation energy (PCE) ratio is used as the correlation plane metric in both filter synthesis and in tests, since it works better than correlation peak height. In tests, we do not assume that the test inputs pose is known (as most prior work does), since pose estimation of SAR objects has a large margin of error; we describe our procedure for proper use of pose estimates in MSTAR recognition. The use of circular versus linear correlations is addressed. We also address the use of multi-look SAR data to improve performance.

Analysis of kernel distortion-invariant filters

Kernel techniques have been used in support vector machines (SVMs), feature spaces, etc. In kernel methods, the wellknown kernel trick is used to implicitly map the input data to a higher-dimensional feature space. If all terms can be written as a kernel function, one can then use data in higher-dimensional space without actually computing the higherdimensional features or knowing the mapping function Φ. In this paper, we address kernel distortion-invariant filters (DIFs). Standard DIFs are synthesized in a linear feature space (in the image or Fourier domain). They are fast since they use FFT-based correlations. If the data is mapped to a higher-dimensional feature space before filter synthesis and before performing correlations, kernel filters result and performance can be improved. Kernel versions of several DIFs (OTF, SDF, and Mace) have been presented in prior work. However, several key issues were ignored in all prior work. These include : the unrealistic assumption of centered data in tests, the significantly larger storage and on-line computation time required, and the proper type of energy minimization in filter synthesis to reduce false peaks is necessary when the filters are applied to target scenes and has yet to be done. In addition, prior kernel DIF work used test set data to select the value of the kernel parameter. In this paper, we analyze these issues, present supporting test results on two face databases, and present several improvements to prior kernel DIF work.

Face recognition with illumination and pose variations using MINACE filters

This paper presents the status of our present CMU face recognition work. We first present a face recognition system that functions in the presence of illumination variations. We then present initial results when pose variations are also considered. A separate minimum noise and correlation energy (MINACE) filter is synthesized for each person. Our concern is face identification and impostor (non-database face) rejection. Most prior face identification did not address impostor rejection. We also present results for face verification with impostor rejection. The MINACE parameter c trades-off distortion-tolerance (recognition) versus discrimination (impostor rejection) performance. We use an automated filter-synthesis algorithm to select c and to synthesize the MINACE filter for each person using a training set of images of that person and a validation set of a few faces of other persons; this synthesis ensures both good recognition and impostor rejection performance. No impostor data is present in the training or validation sets. The peak-tocorrelation energy ratio (PCE) metric is used as the match-score in both the filter-synthesis and test stages and we show that it is better than use of the correlation peak value. We use circular correlations in filter synthesis and in tests, since such filters require one-fourth the storage space and similarly fewer on-line correlation calculations compared to the use of linear correlation filters. All training set images are registered (aligned) using the coordinates of several facial landmarks to remove scale variations and tilt bias. We also discuss the proper handling of pose variations by either pose estimation or by transforming the test input to all reference poses. Our face recognition system is evaluated using images from the CMU Pose, Illumination, and Expression (PIE) database. The same set of MINACE filters and impostor faces are used to evaluate the performance of the face identification and verification systems.

Distortion-invariant kernel filters for general pattern recognition

We note several key general pattern recognition (GPR) issues that have been ignored in all prior distortion-invariant kernel filter (kernel DIF) work. These include: the unrealistic assumption of centered test data, the lack of a fast FFTbased on-line implementation, the significantly larger storage and on-line computation requirements, incorrect formulation of the kernel filter in the FT domain, incorrect formulation of prior image-domain kernel SDF and Mace filters, and the unrealistic use of test set data for parameter selection. We present several improvements to prior kernel DIF work. Our primary objective is to examine the viability of kernel DIFs for GPR and automatic target recognition (ATR) applications (where the location of the object in the test input is not known). Thus, in this paper, we apply our improved kernel DIFs to CAD ATR data. We address range and full 360° aspect view variations; we also address rejection of unseen confuser objects and clutter. We use training and validation set data (not test set data) to select the kernel parameter. We show that kernel filters (higher-order features) can improve classification and confuser rejection performance. We consider only kernel SDF filters, since their on-line computation requirements are reasonable; we present test results for both polynomial and Gaussian kernels. The main purposes of this paper are to: note issues of importance ignored in all prior kernel DIF work, detail how to properly perform energy minimization in kernel DIFs, show that kernel SDF filters can correct errors for ATR data, and compare the performance of kernel SDF filters and standard Minace DIFs. We also introduce our new Minace-preprocessed kernel SDF filter.

Minace filter tests on the Comanche IR database

This paper presents our IR automatic target recognition (ATR) work on the Comanche database using the minimum noise and correlation energy (MINACE) distortion-invariant filter (DIF). The Comanche database contains real IR data of eight targets with aspect view and thermal state variations. We consider recognition of six of these targets and we consider rejecting two targets (confusers) and clutter. To handle the full 360° range of aspect view in Comanche data, we use a set of Minace filters for each object; each filter should recognize the object in some angular range. We use our autoMinace algorithm that uses a training and a validation set to select the Minace filter parameter c (which selects emphasis on recognition or discrimination) and to select the training set images to be included in the filter, so that the filter can achieve both good recognition and good confuser and clutter rejection performance. No confuser, clutter, or test set data are present in the training or the validation set. Use of the peak-to-correlation energy (PCE) ratio is found to perform better than the use of the correlation peak height metric. The use of circular versus linear correlations is addressed; circular correlations require less storage and fewer online computations and are thus preferable.

Automated distortion-invariant filter synthesis and training set selection (auto-Minace)

The minimum noise and correlation energy (MINACE) distortion-invariant filter (DIF) finds use in several applications such as face recognition, automatic target recognition (ATR), etc., in which one considers both true-class object classification and rejection of non-database objects (that are labeled as impostors in face recognition, and confusers in ATR). To solve the classification/rejection problem, we use at least one Minace filter per object class to be recognized. A separate Minace filter or a set of Minace filters is synthesized for each object class. The Minace parameter c trades-off distortion-tolerance (recognition) versus discrimination (impostor/confuser rejection) performance. We present our automated Minace filter-synthesis algorithm (auto-Minace) that selects the training set images to be included in the filter and selects the filter parameter c, so that the filter can achieve both good recognition and impostor/confuser and clutter rejection performance; this is achieved using a training and validation set. No impostor/confuser, clutter or test set data is present in the training or validation sets. The peak-to-correlation energy (PCE) ratio is used as the correlation plane metric in both filter synthesis and in tests, since it gives better results than use of the correlation peak value. We also address the use of the Minace filters in detection applications where the filter template is much smaller than the target scene. The use of circular versus linear correlations are addressed, circular correlations require less storage and fewer online computations.

Automated synthesis of distortion-invariant filters: AutoMinace

This paper presents our automated filter-synthesis algorithm for the minimum noise and correlation energy (MINACE) distortion-invariant filter (DIF). We discuss use of this autoMinace filter in face recognition and automatic target recognition (ATR), in which we consider both true-class object classification and rejection of non-database objects (impostors in face recognition and confusers in ATR). We use at least one Minace filter per object class to be recognized; a separate Minace filter or a set of Minace filters is synthesized for each object class. The Minace parameter c trades-off distortion-tolerance (recognition) versus discrimination (impostor/confuser/clutter rejection) performance. Our automated Minace filter-synthesis algorithm (autoMinace) automatically selects the Minace filter parameter c and selects the training set images to be included in the filter, so that we achieve both good recognition and good impostor/confuser and clutter rejection performance; this is achieved using a training and validation set. No impostor/confuser, clutter or test set data is present in the training or validation sets. Use of the peak-to-correlation energy (PCE) ratio is found to perform better than the correlation peak height metric. The use of circular versus linear correlations is addressed; circular correlations require less storage and fewer online computations and are thus preferable. Representative test results for three different databases - visual face, IR ATR, and SAR ATR - are presented. We also discuss an efficient implementation of Minace filters for detection applications, where the filter template is much smaller than the input target scene.

Face verification and rejection with illumination variations using MINACE filters

A face verification system based on the use of a minimum noise and average correlation energy (MINACE) filter for each person is presented that functions with illumination variations present. A separate filter is used for each person; it is a combination of different training images of only that person. The system is tested using both unregistered and registered images from the CMU Pose, Illumination and Expression (PIE) database. The number of correct (PC) and the number of false alarm (PFA) scores are compared for the two cases. Rather than using the same parameters for the filter of each person, an automated iterative filter training and synthesis method is used. A validation set of several other faces is used to achieve parameter selection for good rejection performance. For filter-evaluation, all filters are tested against all images, but the same peak threshold is used for each filter to determine verification and rejection.