Model-based image analysis for forensic shoe print recognition

This thesis is about automated forensic shoe print recognition. Recognizing a shoe print
in an image is an inherently difficult task. Shoe prints vary in their pose, shape and
appearance. They are surrounded and partially occluded by other objects and may
be left on a wide range of diverse surfaces. We propose to formulate this task in a
model-based image analysis framework.
Our framework is based on the Active Basis Model. A shoe print is represented as
hierarchical composition of basis filters. The individual filters encode local information
about the geometry and appearance of the shoe print pattern. The hierarchical com-
position encodes mid- and long-range geometric properties of the object. A statistical
distribution is imposed on the parameters of this representation, in order to account for
the variation in a shoe print‘s geometry and appearance.
Our work extends the Active Basis Model in various ways, in order to make it robustly
applicable to the analysis of shoe print images. We propose an algorithm that automat-
ically infers an efficient hierarchical dependency structure between the basis filters. The
learned hierarchical dependencies are beneficial for our further extensions, while at the
same time permitting an efficient optimization process. We introduce an occlusion model
and propose to leverage the hierarchical dependencies to integrate contextual informa-
tion efficiently into the reasoning process about occlusions. Finally, we study the effect
of the basis filter on the discrimination of the object from the background. In this con-
text, we highlight the role of the hierarchical model structure in terms of combining the
locally ambiguous filter response into a sophisticated discriminator.
The main contribution of this work is a model-based image analysis framework which
represents a planar object‘s variation in shape and appearance, it‘s partial occlusion as
well as background clutter. The model parameters are optimized jointly in an efficient
optimization scheme. Our extensions to the Active Basis Model lead to an improved
discriminative ability and permit coherent occlusions and hierarchical deformations. The
experimental results demonstrate a new state of the art performance at the task of
forensic shoe print recognition.