Object Recognition using Moments of the Signature Histogram
MetadataShow full metadata
The amount of digital information generated each day is increasing at a very high rate. Our ability to understand and make sense of such large amounts of unstructured data depends on efficient and reliable methods to automate analysis and classification. Still and motion imagery make up a large part of this ever expanding digital universe and therefore methods that target images and video data are increasingly important. The field of image processing has grown to meet this demand and, in particular, techniques for recognizing objects are playing a central role in this field. Digital image processing is a continuum of processes and procedures. This paper is concerned with mid-level image processing, involving segmentation of an image into regions or objects, description of those objects, as well as recognition, and classification of those objects. Specifically, techniques and methods to recognize individual objects in images are investigated. The goal of this thesis is to address the problem of analyzing and matching image objects. To achieve this goal, the use of statistical moments of the object signature is investigated. An object signature is derived by taking the Euclidean distance from the centroid of the object to every pixel on the boundary of the object. A relative frequency histogram is constructed from the object signature and then used to approximate a probability density function for the signature. Statistical moments are then applied to the histogram to generate a novel set of descriptors that are invariant to rotation, translation, and scaling. Existing techniques that utilize moments of the entire image are examined along with moments applied to just the object contour. Additionally, the use of two-dimensional Fourier Descriptors applied to the object contour are considered as well as one- dimensional Fourier Descriptors applied to the object signature. Finally, moments applied directly to the object signature are investigated. Experiments are performed to evaluate and compare these techniques with the method introduced in this work. Recognition accuracy as well as the quality of recognition are used to differentiate between the various techniques. The results of the experiments show the method introduced in this work, statistical moments of the histogram of the object signature, proves to be a viable alternative to the other methods discussed. In particular, since only the center bin-values of the constructed histogram are used to calculate moments, the computational costs are orders of magnitude smaller than the computational cost of other methods considered in this thesis. In addition, the effect of binning the data when constructing the histogram compensates for noise introduced by scaling and rotation, resulting in an improvement in the quality of recognition over several of the other methods investigated.