Minace filter infrared target tracking, recognition, and rejection tests with aspect view, depression angle, and scale variations

Abstract

We examine the sensitivity of minimum noise and correlation energy (MINACE) filters to three different types of distortion variations (aspect view, depression angle, and scale) that are typically present in infrared (IR) imagery used for automatic target recognition (ATR) and tracking applications. Prior DIF (distortion-invariant filter) ATR work has addressed at most two simultaneous variations -aspect view and depression angle variations for SAR data, and aspect view and thermal state variations for IR data. No prior Minace ATR work has addressed scale variations. In our tests, we consider all three simultaneous variations -aspect view, depression angle, and scale. This is new. Our goal is to determine if one Minace filter per object can handle full 360° aspect view variations and can handle small depression angle variations, and to determine the range of scales that one Minace filter per object can handle after training on data at one or more scales. This determines when new Minace filters are needed in an image closing sequence. In all cases, shifts of the target test inputs are considered. We use our autoMinace algorithm that automates selection of the Minace filter parameter c and the training set images to be included in the filter. We also consider rejection of unseen confuser objects and clutter. No confuser, clutter, or test set data are present in the training or the validation set. We present test results using both real and CAD IR data.

Kernel synthetic discriminant function (SDF) filters for fast object recognition

“…pixels 1 and 2) in two training images i and j, i.e., x i [1] x j [1] x i [2] x j [2], etc. Equation (15) then becomes…”

“…We now write the first term (x i [1] x j [1] + x i [2] x j [2] + …) 2 as the sum of two terms: one contains the square of the product of the same pixel in two training images i and j, i.e., (x i [1] x j [1]) 2 , etc., and the other contains the cross-pixel product term between two different pixels (e.g. pixels 1 and 2) in two training images i and j, i.e., x i [1] x j [1] x i [2] x j [2], etc.…”

“…i.e., a kernel is first calculated only for pixel m = 1 in training images i and j (this is K(x i [1], x j [1]); this is repeated for all m = 1…d pixels and these are summed up for all d pixels. We refer to this version as the pixel-based kernel.…”

“…We emphasize one type of classifier, the distortion-invariant filter (DIF), for fast object recognition, since it can be applied for different shifts using the fast Fourier transform (FFT). A single DIF [1] handles all aspect-view distortions, a range of scale distortions, and a ±2° range of depression-angle distortions. DIFs all use a combination of training set images that are representative of the expected distortions in the test set.…”

A new SVM for scale, aspect, and depression angle tolerant IR object recognition

“…However, most of them did not consider rejection of unseen confuser objects when designing correlation filters as summarized in [2] (SAR), [3] (IR), and elsewhere; thus much correlation filter work did not address such a realistic and important issue in ATR. Other comparisons of ATR classifiers [4,5] show large false alarm rates for various standard classifiers (such as SVMs), while other comparisons consider classifiers trained on the false class to be rejected [6,7], which is not realistic.…”