Invariance in pattern recognition: application to line images

摘要

The general aim of pattern recognition is to find algorithms (operators or programs) which map a set of data-representations into a set of results-interpretations. The main difficulty is the algorithmic complexity of pattern recognition processes, and thus the intractability of pattern recognition problems. This is overcome by decomposing pattern recognition problems into sub-problems, so that tractable but suboptimal algorithms can be used. It is shown how two concepts can be used as effective guidelines in determining simultaneously features and their corresponding operators. These two concepts are: invariance to transformations of the features and operators of the objects, and variable precision, particularly of the location of features. These principles are implemented and demonstrated at different levels of recognition: in the application used here the first is local and the second global. The first mapping is from pixels to feature detection. Unitary operators or special operators are used on line images; their invariance to transformation is demonstrated. The location of features is undetermined inside the window. The second mapping is from the features to the identification of simple objects, such as printed letters. Invariant descriptions are used: chains of partially ordered features obtained by scans of the image with variable precision. Applications are shown to demonstrate the recognition of omnifont letters (without learning), previously seen manuscript words, and more generally the compression, analysis and recognition of images from both straight or curved lines.