Master of Science in Industrial Engineering (MSIE)
One of the major hurdles in semantic image classification is that only low-level features can be reliably extracted from images as opposed to higher level features (objects present in the scene and their inter-relationships). The main challenge lies in grouping images into semantically meaningful categories based on the available low-level visual features of the images. It is important that we have a classification method that will handle a complex image dataset with not so well defined boundaries between clusters. Learning Vector Quantization (LVQ) neural networks offer a great deal of robustness in clustering complex datasets. This study presents a semantic image classification using LVQ neural network that uses low level texture, shape, and color features that are extracted from images from rural and urban domains using the Box Counting Dimension method (Peitgen et al. 1992), Fast Fourier Transformation and HSV color space. The performance measures precision and recall were calculated while using various ranges of input parameters such as learning rate, iterations, number of hidden neurons for the LVQ network. The study also tested for the feature robustness for image object orientation (rotation and position) and image size. Our method was compared against the method given in Prabhakar et al, 2002. The precision and recall while using various combination of texture, shape, and color features for our method was between .68 and .88, and 0.64 and .90 respectively compared against the precision and recall (for our image data set) of 0.59 and .62 for the method given by Prabhakar et al., 2002.
Document Availability at the Time of Submission
Release the entire work immediately for access worldwide.
Thulasiraman, Prakash, "Semantic classification of rural and urban images using learning vector quantization" (2005). LSU Master's Theses. 1909.
Gerald M. Knapp