Feature Extraction from Images of Mass Disaster Victims

Earthquakes, hurricanes, terrorist attacks, and other such events can cause tremendous harm to people and infrastructure, often leaving those outside the areas impacted with little or no information on the state of their friends and relatives who may have been affected. Our goal is to automatically process images of patients taken as part of the intake process at emergency medical care centers to extract searchable, text-based descriptors of patients that can be accessed remotely (e.g., via the web) to facilitate identification of victims. Achieving this goal requires finding the person in the image, and in particular their face, to drive the process of extracting features such as hair color and style, clothing color and style, the presence or absence of glasses, and other such features. Face detection has been extensively studied for clear, frontal face images. Images under consideration here are characterized by their noisy, non-frontal, blurry nature. We studied the performance of standard existing face detection algorithms and observed that they are exceptionally poor for this domain of images. We developed an ensemble-based face detection method that is a combination of robust skin detection and simple pattern matching face detection techniques. Also, a template-based eye detection method is pro- posed that is used as a feedback system to improve the accuracy of the face detection technique. Finally, we explored eyeglasses detection using color histograms to illustrate feature extraction using new face detection and eye detection techniques