Ph.D Dissertation Defense “AN AUTOENCODER-BASED IMAGE DESCRIPTOR FOR IMAGE MATCHING AND RETRIEVAL” by Chenyang Zhao

ABSTRACT:

Local image features are needed in many computer vision applications. For this purpose, a large number of point detectors and descriptors have been developed throughout the years. However, creation of effective descriptors is still a topic of research. The Scale Invariant Feature Transform (SIFT) proposed by David Lowe is an example of a widely used image descriptor in image analysis and image retrieval. SIFT first detects interest points in an image based on a scale-space analysis. It then creates a descriptor for each detected point, encoding gradient information in a patch centered at the point. SIFT descriptors are scale and rotation invariant and provide a high matching rate; however, they are computationally too slow to be useful in many image analysis and retrieval applications.

Autoencoder is an effective computational method for representation learning. In this dissertation it is used to construct a low-dimensional representation for a high-dimensional data while preserving structural information within the data. In many computer vision applications, a high dimensional image data implies a high computational cost. The main motivation in this work is to improve the speed of image descriptors significantly without reducing their match ratings noticeably. A new descriptor is designed that is based on the autoencoder concept. The proposed descriptor can reduce the size and complexity of a descriptor significantly, considerably reducing the time required to find an object of interest in an image or retrieve a desired image from a database.

Ph.D. Committee: Drs. Arthur Goshtasby, Advisor, Thomas Wischgoll, Jack Jean, and Caroline Cao, BIE