PhD Oral Exam - Tamara Finide Dittimi, Computer Science

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

Banknote recognition and medical image analysis have been the foci of image processing and pattern recognition research. As counterfeiters have taken advantage of the innovation in print media technologies for reproducing fake monies, hence the need to design systems which can reassure and protect citizens of the authenticity of banknotes in circulation. Similarly, many physicians must interpret medical images. But image analysis by humans is susceptible to error due to wide variations across interpreters, lethargy, and human subjectivity. Computer-aided diagnosis is vital to improvements in medical analysis, as they facilitate the identification of findings that need treatment and assist the expert’s workflow. Thus, this thesis is organized around three such problems related to Banknote Authentication and Medical Image Diagnosis.

In our first research problem, we proposed a new banknote recognition approach that classifies the principal components of extracted HOG features. We further experimented on computing HOG descriptors from cells created from image patch vertices of SURF points and designed a feature reduction approach based on a high correlation and low variance filter.

In our second research problem, we developed a mobile app for banknote identification and counterfeit detection using the Unity 3D software and evaluated its performance based on a Cascaded Ensemble approach. The algorithm was then extended to a client-server architecture using extracted SIFT and SURF features reduced using Bag of Words and high correlation-based HOG vectors.

In our third research problem, experiments were conducted on a pre-trained mobile app for medical image diagnosis using three convolutional layers with an Ensemble Classifier comprising PCA and bagging of five base learners. Also, we implemented a Bidirectional Generative Adversarial Network to mitigate the effect of the Binary Cross Entropy loss based on a Deep Convolutional Generative Adversarial Network as the generator and encoder with Capsule Network as the discriminator while experimenting on images with random composition and translation inferences. Lastly, we proposed a variant of the Single Image Super-resolution for medical analysis by redesigning the Super Resolution Generative Adversarial Network to increase the Peak Signal to Noise Ratio during image reconstruction by incorporating a loss function based on the mean square error of pixel space and Super Resolution Convolutional Neural Network layers.