Classification of ground penetrating radar images using histogram of oriented gradients and support vector mechine

Ground Penetrating Radar or generally known as GPR is an important and popular method in subsurface imaging due to its non-destructive nature. GPR data interpretation requires expertise from human operator which is a time consuming and costly task as the data amount can be enormously large. In this...

Full description

Saved in:
Bibliographic Details
Main Author: Lee, Kher Li
Format: Thesis
Language:English
Published: 2016
Subjects:
Online Access:http://eprints.utm.my/id/eprint/60704/1/LeeKherLiMFKE2016.pdf
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Ground Penetrating Radar or generally known as GPR is an important and popular method in subsurface imaging due to its non-destructive nature. GPR data interpretation requires expertise from human operator which is a time consuming and costly task as the data amount can be enormously large. In this study, a framework that pairs up Histogram of Oriented Gradients (HOG) and Support Vector Machine (SVM) is proposed to detect subsurface targets in GPR data automatically. HOG feature descriptors are extracted by characterizing the target appearance and shape from hyperbolic signatures that appear in GPR images. Extracted feature descriptors are then sent to SVM for classification. Contribution of this research includes designing the best SVM classifier model by considering the best kernel and its optimized parameter settings. The proposed algorithm is compared to the most commonly used approach (Hough Transform) to evaluate its performance. In this research, the data sets consist of images that are collected using different GPR system models. Despite having limited sample images for training, the proposed method managed to detect hyperbolic signatures in GPR images. SVM classifier with probabilistic estimation model shows better performance for its flexibility in decision making using confidence level while SVM without probabilistic estimation model shows high false positive rate of more than 50%. Moreover, results from the experiments have also shown that the proposed method is able to produce higher detection rate with a much lower false positive rate than that of Hough Transform. The accuracy of target detection using the proposed method records an average detection rate of 89.40% and 7.38% of false positive rate for all the data sets used in this research. Apart from the improved performance, the proposed method also offers flexibility to control detection tasks through an adjustment on the probabilistic estimation model.