A hybrid pattern recognition paradigm using moment invariants and polynomial networks for segmenting objects in multi-spectral imagery.

摘要

The image understanding task of object classification can be described as segmenting objects from their background, extracting object features, and assigning class definitions to these features. Successful object classification is highly dependent upon initial segmentation of an object from its background. For complex, real-world imaging applications, this task is extremely challenging and critical to the success of the recognition system. Traditional object segmentation techniques rely heavily upon noise removal during pre-processing, and subsequently employs one of two image-level strategies: histogram analysis or region growing. Because effective noise removal strategies are difficult to develop for actual imagery, the success of the overall classification strategy often falls short of requirements. Therefore, alternate methods are required for object segmentation.; An alternate approach is to determine Target/Non-Target status of image regions at the pixel level. In this manner, noise removal and object segmentation are performed in a single process, taking advantage of the large amount of information contained in present-day, multi-spectral imagery. The key issues associated with this approach are proper determination of a pixel information representation and choice of an information fusion algorithm to process pixel-level information. These questions were addressed during this dissertation research.; The goal of this research was to design, develop, and demonstrate an object segmentation paradigm that is robust in the face of noise, clutter, and other adverse, real-world conditions. To achieve this objective, the research integrated multi-spectral imagery (co-registered laser radar and thermal) of real-world scenes, a pixel classification strategy for object segmentation, moment invariants feature extraction algorithms for pixel characterization, and polynomial networks for feature processing. This approach is unique in that it is the first to integrate these advanced image understanding technologies.; To validate the proposed approach, the research compared the utility provided by moment invariants with conventional pixel-level features, heuristically assessed segmentation results, and determined the processing requirements for an operational implementation of the resulting object segmentation methodology.