Feature generation for quantification of visual similarity

Abstract

Contributions of this thesis include an efficient implementation of OBVIS and SDA, guidelines for setting layer sizes in SDA from an implementation efficiency perspective, an analysis of the strength and weakness of OBVIS and SDA, and, finally, a system design that combines OBVIS and SDA to learn a good feature generation process using ground truth data from human studies.

The complex nature of visual similarity makes it extremely difficult to hand code a set of good features that incorporate all of the important aspects for all images. This thesis work shows that machine learning techniques can be used to generate statistically optimal low dimensional features that work well with calculating similarity using Euclidean distance between feature representation of images. Specifically, a Stacked Denoising Autoencoder (SDA) was used to train a deep neural network to learn a set of important features from the Amsterdam Library of Object Images. Theses features generated by SDA were compared with those generated using OBVIS, a feature generation algorithm developed specifically for human visual similarity comparison. The results indicated that features learned by SDA, a generic representation learning approach, outperformed the features generated by OBVIS, a method coded with domain specific knowledge.