Modeling automated detection of children in images
During a child pornography investigation, law enforcement officers are required to manually sort through potentially hundreds of thousands of images found on seized digital media. This process is error prone, time-consuming, and places considerable strain on department resources as well as the investigators themselves. This thesis examines the use of automated facial feature detection to estimate the age of a human in an image. The ability to automate the detection between an adult and a child in a sexually explicit digital image will assist law enforcement agencies combating the production and trafficking of child pornography. The proposed method uses open source Haar classifiers to extract facial feature information from images, and predict a subjects age based on anthropometric computations. Classification, regression, and hybrid approaches to machine learning are used to train and test age-estimation models. Manually-plotted facial landmarks are used as ground truth labels to measure the accuracy of 1) automated feature detection of the eyes, nose and mouth, and 2) age-estimation based on the extracted features. Currently, this methodology has been implemented to distinguish between classes of children and adults with an accuracy of approximately 70%.
Medical imaging|Computer science
"Modeling automated detection of children in images"
Dissertations and Master's Theses (Campus Access).