PhD Dissertation Defense: Cristopher Flagg
A Zoom link for the defense will be distributed through the department mailing lists, or can be obtained by contacting Cris directly.
Title: “Image Transformation of Line Drawing for Improved Retrieval”
Initial data representations, such as images, provide basic visual information about a document collection but contain no deeper relationships between elements within the images or elements within a related collection of images. Transformation of initial image data to an intermediate representation allows neural network models to encode inter-image and intra-image relationships to improve retrieval.
Direct comparison of images and a study of the interrelationship of multiview images derived from 3D model repositories shows information encoded in the inter-image relationships provides stronger retrieval results than is possible with the individual images. This inter-image relationship may be further leveraged to reconstruct the original 3D model. Encoding the 3D model into a vector representation provides more accurate retrieval than previously explored multiview image retrieval methods.
Trademarks are typically represented as line drawings containing a collection of multiple design elements. The definitions of trademark similarity are specified by both rules set by the Trademark office and precedents set in the courts. To allow flexibility in trademark retrieval, the intra-image features present in trademarks may be used within a search results reranking framework to improve retrieval. Reranking allows retrieval results to conform to intra-image dominant image features.