An adaptive learning approach to acquiring geometry concepts in images

The ability to classify low-level sub-symbolic and symbolic concepts in images is fundamental to the general problem of intelligent machine vision and is often necessary for specific complex industrial and professional applications. The ALISA Texture Module (ATM) using statistical learning techniques with general-purpose features has demonstrated robust performance for the classification of sub-symbolic (i.e., texture-based) concepts in images. For the classification of symbolic concepts in images, where the relative spatial arrangement of regions of different textures is important, the state-of-the-art is currently limited to model-based approaches, which are often brittle and require a large number of parametric assumptions about the problem domain.;The AGM is trained with a few sets of exemplars of texture class maps, each set representing a specific symbolic geometry class. Once trained, the AGM can be used to recognize and classify arbitrary spatial arrangements of the learned geometric structures in texture class maps that have not been used for training. The output of the AGM is a class map that identifies the location and class of each of the learned geometry classes in each input texture class map.;The classification performance of the AGM was validated with two kinds of geometry classes: canonical and secular. Canonical classes (e.g., HORIZONTAL, CURVED, T-JUNCTION) have universal applicability. Secular classes (e.g., types of postage stamps, graphic icons) are application-specific. For both canonical and secular classes, the evaluation of classification performance of the AGM in the presence of noise and under variations of translation, rotation, and scale conclusively demonstrates the ability of the AGM to use a general-purpose feature set to learn and classify a variety of geometry concepts.;By accepting texture class maps as input, the AGM provides a seamless transition from sub-symbolic to symbolic processing, Thus, this research represents a first step toward applying statistical learning techniques to symbolic image-understanding problems. As a logical next step, the ALISA Shape Module, which is currently under development, uses the AGM geometry class maps as input and generates output class maps that identify the location and class of canonical and/or secular shapes that have been learned during training.;In contrast, the adaptive learning engine developed in this research, known as the ALISA Geometry Module (AGM), extracts a set of general-purpose features from segmented texture class maps, in which pixel values represent texture class labels rather than intensity values. The adaptive learning paradigm used for the AGM is Bock's Collective Learning Systems (CLS) Theory, a statistical method that acquires histograms of multi-dimensional feature spaces through supervised training to estimate the discrete probability density function of this space. Unlike most model-based approaches, the AGM uses the same set of features to capture all geometry concepts. The primary research hypothesis of this work is that the AGM, using these general purpose features, can be used to classify a wide variety of geometry concepts, including both those that can and cannot be described parametrically.;The major contribution of this work is the design of a general-purpose feature-vector that allows statistical learning techniques to be applied to segmented images, known as class maps, for the purpose of classifying low-level symbolic geometry concepts. Although the system has not been optimized for speed, the AGM is able to classify a single class map within a few seconds using a desktop computer system.