Jack Sklansky

A note on genetic algorithms for large-scale feature selection

Abstract We introduce the use of genetic algorithms (GA) for the selection of features in the design of automatic pattern classifiers. Our preliminary results suggest that GA is a powerful means of reducing the time for finding near-optimal subsets of features from large sets.

Comparison of algorithms that select features for pattern classifiers

Abstract A comparative study of algorithms for large-scale feature selection (where the number of features is over 50) is carried out. In the study, the goodness of a feature subset is measured by leave-one-out correct-classification rate of a nearest-neighbor (1-NN) classifier and many practical problems are used. A unified way is given to compare algorithms having dissimilar objectives. Based on the results of many experiments, we give guidelines for the use of feature selection algorithms. Especially, it is shown that sequential floating search methods are suitable for small- and medium-scale problems and genetic algorithms are suitable for large-scale problems.

Finding circles by an array of accumulators

We describe an efficient procedure for detecting approximate circles and approximately circular arcs of varying gray levels in an edge-enhanced digitized picture. This procedure is an extension and improvement of the circle-finding concept sketched by Duda and Hart [2] as an extension of the Hough straight-line finder [6].

ON AUTOMATIC FEATURE SELECTION

We review recent research on methods for selecting features for multidimensional pattern classification. These methods include nonmonotonicity-tolerant branch-and-bound search and beam search. We describe the potential benefits of Monte Carlo approaches such as simulated annealing and genetic algorithms. We compare these methods to facilitate the planning of future research on feature selection.

Fast polygonal approximation of digitized curves

Abstract We describe a new technique for fast “scan-along” computation of piecewise linear approximations of digital curves in 2-space. Our method is derived from earlier work on the theory of minimum-perimeter polygonal approximations of digitized closed curves. We demonstrate the specialization of this technique to the case where the error is measured as the largest Hausdorff-Euclidean distance between the approximation and the given digitized curve. We illustrate the application of this procedure to the boundaries of the images of a lung and a rib in chest radiographs.

The use of Markov Random Fields as models of texture

Abstract We propose Markov Random Fields (MRFs) as probabilistic models of digital image texture where a textured region is viewed as a finite sample of a two-dimensional random process describable by its statistical parameters. MRFs are multidimensional generalizations of Markov chains defined in terms of conditional probabilities associated with spatial neighborhoods. We present an algorithm that generates an MRF on a finite toroidal square lattice from an independent identically distributed (i.i.d.) array of random variables and a given set of independent real-valued statistical parameters. The parametric specification of a consistent collection of MRF conditional probabilities is a general result known as the MRF-Gibbs Random Field (GRF) equivalence. The MRF statistical parameters control the size and directionality of the clusters of adjacent similar pixels which are basic to texture discrimination and thus seem to constitute an efficient model of texture. In the last part of this paper we outline an MRF parameter estimation method and goodness of fit statistical tests applicable to MRF models for a given unknown digital image texture on a finite toroidal square lattice. The estimated parameters may be used as basic features in texture classification. Alternatively these parameters may be used in conjunction with the MRF generation algorithm as a powerful data compression scheme.

Estimating the 3D skeletons and transverse areas of coronary arteries from biplane angiograms

A method for estimating the three-dimensional (3D) skeletons and transverse areas of the lumens of coronary arteries from digital X-ray angiograms is described. The method is based on the use of a 3D generalized cylinder (GC) consisting of a series of 3D elliptical disks transverse to and centered on a 3D skeleton (medial axis) of the coronary arteries. The estimates of the transverse areas are based on a nonlinear least-squares-error estimation technique described by D.W. Marquardt (1963). This method exploits densitometric profiles, boundary estimates, and the orientation of the arterial skeleton in 3-space and includes an automatic artery tracking procedure. It applies an adaptive window to the densitometric profile data that are used in the parameter estimation. Preliminary experimental tests of the procedure on angiograms of in vivo human coronaries and on synthetic images yield encouraging results.

Finding the convex hull of a simple polygon

We describe a new algorithm for finding the convex hull of any simple polygon specified by a sequence of m vertices. An earlier convex hull finder of ours is limited to polygons which remain simple (i.e., nonselfintersecting) when locally non-convex vertices are removed. In this paper we amend our earlier algorithm so that it finds with complexity O(m) the convex hull of any simple polygon, while retaining much of the simplicity of the earlier algorithm.

Recognition of convex blobs

Abstract Because of the discrete nature of the memory and logic of a digital computer, a digital computer “sees” pictures in cellular form, each cell containing a number that represents the density of the viewed object at that cell. In particular, when the picture is binary, each cell holds a 1 or 0, depending on whether or not the viewed object is projected onto that cell. The convexity of cellular blobs—a restricted class of binary cellular figures—is discussed and defined in terms of the continuous blobs of which the cellular blobs are images. The elements of a theory of convex cellular blobs are given. As an application of the theory, the use of the “minimum-perimeter polygon” in an algorithm for testing the convexity of cellular blobs on a rectangular mosaic is described.

The Detection and Segmentation of Blobs in Infrared Images

A computer procedure for detecting and finding the boundaries of blobs in noisy infrared images is described. Our evaluation of this procedure on a data base of 81 targets, 34 for design and 47 for test, resulted in only two false negatives (missed targets) and no false detections. Our procedure consists of an intensity normalizer, a dc notch filter, an edge detector, a spoke filter, a gradient-guided segmenter, an extractor of the standard deviation of the gray level in each blob, an extractor of the fraction of intense edge elements along the boundary of each blob, and a three-nearest-neighbor classifier. Among these processes, the spoke filter and the gradient-guided segmenter are new. Both of them contribute strongly to the effectiveness of our procedure. The spoke filter is sensitive to a wide variety of shapes of blobs within a specified range of sizes. The gradient-guided segmenter exploits the noise immunity of the direction of the digital gradient to find a best threshold for segmenting each detected blob.