Abstract
Subspace classifiers have been around for a long time, beginning with feature selection, which in essence was a subspace selection technique. This talk will discuss the kind of subspace classifiers that Bledsoe and Browning presented in their 1959 paper and from which there have been a variety of extensions which we will discuss.
The Bledsoe and Browning subspace classifier quantizes measurement space. Each quantized observation tuple corresponds to a cell in measurement space. A collection of subspaces are selected at random. In the original form the subspaces were mutually exclusive. For each class, each cell of a subspace contained a number dependent on the number of observations of the training data that fell into that cell. For each class those numbers were combined in ways not dissimilar to random forests. For a given observation tuple, the class with the highest vote count was selected as the assigned class.
We will discuss a variety of principled extensions of the technique and make some comparisons with Neural Networks.
Research Interests and Bio
High-dimensional space clustering, pattern recognition, knowledge discovery and artificial intelligence
Professor Haralick began his work as one of the principal investigators of the NASA ERTS satellite data doing remote sensing image analysis.
He has made a series of contributions in the field of computer vision. In the high-level vision area, he has worked on inferring 3D geometry from one or more perspective projection views.] He has also identified a variety of vision problems which are special cases of the consistent labeling problem. His papers on consistent labeling, arrangements, relation homomorphism, matching, and tree search translate some specific computer vision problems to the more general combinatorial consistent labeling problem and then discuss the theory of the look-ahead operators that speed up the tree search. The most basic of these is called Forward Checking. This gives a framework for the control structure required in high-level vision problems. He has also extended the forward-checking tree search technique to propositional logic.
In the low-and mid-level areas, Professor Haralick has worked in image texture analysis using spatial gray tone co-occurrence texture features. These features have been used with success on biological cell images, x-ray images, satellite images, aerial images and many other kinds of images taken at small and large scales. In the feature detection area, Professor Haralick has developed the facet model for image processing. The facet model states that many low-level image processing operations can be interpreted relative to what the processing does to the estimated underlying gray tone intensity surface of which the given image is a sampled noisy version. The facet papers develop techniques for edge detection, line detection, noise removal, peak and pit detection, as well as a variety of other topographic gray tone surface features.
Professor Haralick's work in shape analysis and extraction uses the techniques of mathematical morphology. He has developed the morphological sampling theorem which establishes a sound shape/size basis for the focus of attention mechanisms which can process image data in a multiresolution mode, thereby making some of the image feature extraction processes execute more efficiently. He has also developed recursive morphological algorithms for the computation of opening and closing transforms. The recursive algorithms permit all possible sized openings or closings for a given structuring element to be computed in constant time per pixel.] He also developed statistical morphological methodologies for image analysis and noise removal.
In the area of document image understanding, Professor Haralick was responsible for the development of comprehensive ground-truthed databases consisting of over 1500 document images, most in English and some in Japanese. The databases are issued on CD-ROMs and are used all around the world by people developing character recognition methodologies and techniques for document image structural decomposition. He has developed algorithms for document image skew angle estimation, zone delineation, and word and text line bounding box delineation.
In a series of papers, Haralick has helped influence the computer vision community to be more sensitive to the needs of computer vision performance characterization and covariance propagation for without this kind of analysis Computer Vision has no robust theory.
Professor Haralick has contributed to the medical image analysis area particularly working with X-ray ventriculargrams and echocardiography, His papers developed techniques to identify and delineate anatomically accurate boundaries for the left ventricle of the heart.
His most recent work is in the machine learning area, particularly in the manifold clustering of high dimensional data sets, the application of pattern recognition to mathematical combinatorial problems. He is current work is in the learning of knowledge and structure through relation decomposition.
Professor Haralick is a Fellow of IEEE for his contributions in computer vision and image processing and a Fellow of the International Association for Pattern Recognition (IAPR) for his contributions in pattern recognition, image processing, and for service to IAPR. He served as president of IAPR from 1996 to 1998. He has served on the editorial board of "IEEE Transactions on Pattern Analysis and Machine Intelligence" and has been the computer vision area editor for Communications of the ACM and as an associate editor for Computer Vision, Graphics, and Image Processing, The IEEE Transactions on Image Processing and Pattern Recognition. He served on the editorial board of Real Time Imaging and the editorial board of Electronic Imaging.