AN AUTOMATIC BIO-INSPIRED KEY POINT DETECTION AND EXTRACTION STRATEGY FOR CLASSIFICATION OF 2-D AND 3-D IMAGES

Author

Neda Hantehzadeh, Southern Illinois University CarbondaleFollow

Date of Award

8-1-2015

Degree Name

Doctor of Philosophy

Department

Electrical and Computer Engineering

First Advisor

Gupta, Lalit

Second Advisor

Schlesinger, Matthew

Abstract

In this dissertation an optimum key point detection and extraction model is introduced and tested for computer vision applications. In particular, applications of automatic classification of synthetic similar shapes of airplanes and 3-dimensional faces are extensively explored. The model implementation of this study focuses on information captured by Low-level human vision to identify invariant robust feature of faces or shapes for recognition/classification tasks. The model utilizes a modified saliency representation of Itti. The high salient regions are selected using an automatic stochastic process. Each feature on the saliency map in this study consists of a statistical representation of image orientation, intensity and color using two or three-dimensional image information. The main goal of this study is to show the saliency transformation of the images that accounts for detection of robust regions along with the extraction of key point descriptors can lead to high quality and accurate classification performances regardless of geometric and spatial transformation of the images. The model’s performance is evaluated by generating synthetic images of shapes of airplanes that contain random high-density noise, occlusion as an addition to the general geometric transformations such as scale and rotation. Texas 3D face data set is used to analyze the model’s performance for 3D face recognition. It contains variety range of change in illumination, pose and expression. The results demonstrate the model’s invariance to noise, occlusion, scale and rotation for the synthetic shapes and also performs excellent recognition rates considering the changes in pose, illumination and expression for faces. The key to reliable detection and robustness to geometric and spatial transformations lies in the series of image filtering techniques, iterative Difference of Gaussian (DoG) filtering along with the calculation of Haar wavelet descriptors extracted from the saliency values.