Title: A two-stage approach to image segmentation in forensic footwear comparisons
Authors: Adam Pintar - National Institute of Standards and Technology (United States) [presenting]
Steven Lund - National Institute of Standards and Technology (United States)
Rishi Venkatasubramanian - Indian Institute of Science Education and Research Bhopal (India)
Abstract: In forensic footwear comparisons, a shoe print examiner compares an image of a shoe impression from a crime scene (the $Q$ image) to an image of a shoe impression made in a laboratory (the $K$ image) and reaches a conclusion such as identification, exclusion, or inconclusive. Processing the $Q$ image, the $K$ image, or both may aid the comparison. One type of processing, a form of noise reduction, is segmenting the shoe contact from the background in the $Q$ image. However, the complex background observed in many $Q$ images can make segmentation difficult. A two-stage approach to solving the problem is presented. In the first stage, simple linear iterative clustering (SLIC) is used to partition the image into regions. Where regions follow meaningful boundaries between contact and background, they may be used to quickly segment a portion of the $Q$ image. The segmented portion yields training data to feed into a convolutional neural network algorithm for image segmentation known as U-net. The final product is a method to create a single-use neural network for the complete segmentation of one $Q$ image.