A new distortion-invariant correlation filter that truly optimizes parameters is described. Shortcomings of other distortion-invariant correlation filters are noted (selecting the number of training images and the control parameter, worse objective function optimization with more training images, lower clutter energy does not correspond to a better (PFA). A new control parameter (c) definition is used, and new objective functions.
We consider a new detection application for distortion-invariant filters. Several new advances to the MINACE filter are considered. These include: improved object modeling (this provides scene correlation plane peaks near the specified values), improved clutter modeling (no false class training is used), a new trade-off parameter c definition (with the energy of each spectra normalized to one), use of a smaller filter size and a prime factor FFT (this reduces noise effects), zero-mean filters (these allow detection of hot and cold contrast objects), etc. We advance a new peak variance degree of freedom distortion-invariance filter (PVDDF) with many attractive new properties. These advantages include: insurance that the correlation peak values for all distorted objects are close to a given value without requiring that each object have a given correlation peak value (in practice one does not want all distorted object inputs to give the same exact correlation peak value), use of more training images NT without an associated drop in the required threshold and thus better object modeling, and better object function energy E minimization (E reduces as NT increase compared to other filters where E increase as NT increases). This new filter thus achieves both better correlation peak values and better energy minimization (prior filters cannot achieve both goals) by using degrees of freedom.
We consider distortion-invariant filters for detection (i.e. to locate a number of different object classes). For each object, there are two different depression angles, four different contrast ratios, and 18 different aspect views. The objects are present in a variety of different real background clutter. One filer is able to recognize (detect) all 2 X 4 X 18 X 5 equals 720 object versions in clutter with no false alarms using NT equals 36 training set images. The filter uses training objects in a constant background, correlation peak constraints on the NT objects, and minimizes a weighted combination of the correlation plane energy due to the distortion spectrum and a noise spectrum. The new object and noise models used produce this excellent performance with no false class clutter training.
This paper presents a simplified description of coherent and noncoherent optical correlators and derives the correlator models. The advantages of the tolerances in component positioning and quality for the noncoherent correlator are explained in detail. Performance measures for correlator comparison are outlined and quantitative comparison data are provided. An implementation approach for bipolar noncoherent filters is introduced and promising results simulating a limited number of amplitude and phase levels are given. We will show that the performance of noncoherent and coherent correlators are comparable and thus the noncoherent correlator should be more widely used because of its implementation advantages.
This paper concerns the realization of a noncoherent correlator: the positioning tolerances and light budgets of a noncoherent correlator, the effect of the wavelength spread of the noncoherent source, the need for multiplexing noncoherent filters, the best input device (CRT or SLM) to use in a noncoherent correlator, and the light budget for the different noncoherent correlators. We conclude that a laser diode with a diffuser and input SLM laser diode array is the best choice for light source and input device. We demonstrate that a laser diode array can also be used to multiplex CGH's and that this is the preferable architecture.
Advanced version of MINACE (minimum noise and correlation plane energy) filters are applied to detection. Detection involves locating all objects in multiple classes in an image independent of the object class, distortions and contrast differences, i.e. we do not attempt to recognize the object class. This is the first step in scene analysis. A number of false alarms are expected per scene and will be reduced in subsequent levels of the full processor. The advanced MINACE filter concepts we use involve: including a constant background in the training images, training on intermediate contrast ratio objects, use of a new background clutter model and use of orthogonal filter sets.