An EMD based technique for pattern recognition.

摘要

There are a wide variety of approaches in the literature for the automatic detection and classification of a desired object from an image. Any detection scheme requires a training database for the object of interest and depending on the application, the definition of object and clutter (object of no interest) may change considerably, they may even become converse of each other. The automatic object recognition system that is developed in this dissertation has four main stages: (I) Selecting good, discriminative and affine invariant interest points from an image, (II) Extracting characteristic features related to the change in brightness and color, at those selected points, (III) Hierarchical clustering of those points based on the extracted stable local features and position, and (IV) Comparing the clusters with the training data-set by using a similarity measure to yield the final classification or detection result.;In this dissertation, the main contributions are (I) devising a novel and efficient algorithm for detecting affine invariant interest points such that, not only no true interest points will be missed but also no false interest points will be detected in the image, (II) combining partitional clustering and divisive (top-down) clustering to formulate a two-step hierarchical clustering for extracting the possible candidate road signs or the region of interests (ROIs), where the partitional clustering of the detected points is performed based on the stable local features, and then points belonging to each partition are reclustered using position feature. (III) utilizing the distortion invariant fringe-adjusted joint transform correlation (JTC) technique for matching the extracted candidate object regions with the existing known reference objects of interest stored in the database.;Using this method, an algorithm has been developed in this dissertation that reliably detects road signs from the natural scenes and yields a very low false hit rate.;Interest points or corners are sparse and robust features of an image. They provide useful information and give important clues for the shape representation. Some of the main applications of interest points are image matching, object recognition, motion detection, tracking, image mosaicing, panorama stitching, 3-D modeling, etc. The first part of this dissertation introduces a novel contour based method for detecting largely affine invariant interest or feature points, which is able to deal with significant affine transformations including large rotations, shearing and scale changes. In the first step, image edges are detected by morphological operators, followed by edge thinning. In the second step, corner or feature points are identified based on the local curvature of the edges. One of the main contributions of this dissertation is the selection of good discriminative feature points from the thinned edges based on the 1-D empirical mode decomposition (EMD).;Road signs have well defined color, shape, size and position, which aid in the detection tasks. Therefore, in order to differentiate road signs from other objects, the distinctive properties of road signs need to be exploited, such as the color distribution and the geometric constraints. The detected discriminative points are first clustered using the expectation maximization (EM) algorithm based on the local region analysis, for example, the brightness and color features of the neighborhood. In the second part of this dissertation, a hierarchical algorithm has been proposed to find successive clusters from the previously established clusters, where points belonging to each initial partition are reclustered using position feature depending on the spread or scattering of those points relative to each other. This proposed two-step hierarchical clustering yields the possible candidate road signs or the region of interests (ROIs).;In the third part of this dissertation, the distortion invariant fringe-adjusted joint transform correlation (JTC) technique determines whether the candidate ROI contains a road sign and, if it does, assigns a type to that sign. For classification, a normalized composite filter is created using a set of distorted reference images for each road sign type. The query image, i.e., the ROI, is then compared to the composite filters corresponding to the reference road sign images in the database, and the reference road sign most similar to the query sign is returned to the user as the final recognition result.;The presented framework provides a novel way to detect a road sign from the natural scenes and the simulation results demonstrate the efficacy of the proposed technique.