Using machine learning algorithms for criminal identification

• Main Author: Ma, Xiaoxue
• Other Authors: Kong Wai-Kin Adams
• Format: Final Year Project
• Language: English
• Published: 2016
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Pattern recognition
• Online Access: http://hdl.handle.net/10356/66785
• Institution: Nanyang Technological University

With increasing popularity of digital cameras and widespread video surveillance in common public areas, more digital images and video clips are collected as resources and evidence for criminal identification. Traditional facial recognition approach does not handle cases where faces are not observable. Therefore, researchers have explored alternative biometrics for criminal identification. Forearm images can be good candidate because forearms are often uncovered in the images especially in tropical zone. However, forearm biometrics such as vein and tattoos can be difficult to obtain when skin area exposed is not large enough, or when body fat concentration is high. Moreover, criminal identification using forearm image still has lower accuracy compared to other methods, and therefore, a more robust and powerful algorithm to perform identification task on forearm images is needed. Medical studies have shown that androgenic hair patterns can be used for identification. Such algorithm has already been developed for lower leg images. Considering similarities between leg images and forearm images, the algorithm can be applied to forearm images. This study applied androgenic hair pattern matching algorithm on forearm images to perform identity matching. In the experiment using 250 test images and 250 training images, the accuracy achieved by the algorithm is 91.6% at its first rank, which is about 10% higher compared to state-of-art identity matching algorithm based on vein pattern extracted from colour images. Additionally, the androgenic hair pattern matching algorithm, implemented in MATLAB, has been accelerated using multiple techniques. The algorithm performs 82.33% faster after acceleration. A Python version of the algorithm is also implemented to provide the opportunity of low-level optimisation and achieve higher scalability at lower cost. A case study has been provided in this report to illustrate the optimisation and scalability. Based on the outcome of this research work, androgenic hair pattern matching algorithm provides significant improvement over existing algorithms on criminal identification task using forearm images.