Context Based Tattoo Image Analysis with Applications in Public Safety

Joonsoo Kim, Purdue University

Abstract

Law enforcement is interested in exploiting tattoos as an information source to identify, track and prevent gang-related activities. In this thesis we examine several aspects of tattoo image analysis and how to extract useful information from tattoo images. There are problems with existing tattoo image systems. For example, many existing tattoo retrieval systems do not use local and global image descriptors robust to deformations caused by ”manually constructed” tattoos on human skin. One other issue is that most tattoo images are manually cropped to remove background clutter from the image before analysis. In this thesis we examined various aspects of a tattoo image retrieval and classification, in particular we investigated segmentation, classification, image matching and retrieval. A tattoo region is segmented using graph-cut tattoo segmentation based on image edges, a skin color model and a visual saliency map. We generate local and global image descriptors for the segmented image based on multiple polar histograms to introduce robustness against various deformations. The multiple polar histograms are combined with SIFT descriptors in the local image descriptor and 2D DFT in the global image descriptor. We then search our tattoo image database and retrieve images similar to the segmented image using an image matching technique based on both descriptors. To improve the image retrieval accuracy, not only we find the pairwise image similarity between an input image and a database image but we also incorporate all the image similarities between the database image and other database images using inductive matching. For the tattoo image classification, sparse codes based on dense SIFT descriptors are generated. The sparse codes are then combined with a spatial pyramid feature pooling to incorporate the spatial distribution of the sparse codes. A spatial pyramid alignment method is additionally used to improve the image classification accuracy. These methods are evaluated on datasets that were collected from the Indiana State Police, eviltattoo.com, and the NIST tattoo challenge dataset.

Degree

Ph.D.

Advisors

Delp, Purdue University.

Subject Area

Computer Engineering|Electrical engineering|Criminology

Off-Campus Purdue Users:
To access this dissertation, please log in to our
proxy server
.

Share

COinS