An Improved graph mining tool and its application to object detection in remote sensing

Aktaş, Ümit Ruşen
In many graph-based data mining tools, the use of numeric values as attributes in graphs is very limited. Most algorithms require pre-processing of the attributes, which often involves discretization into bins and embedding group names in the input graph(s). In this thesis, we tackle this problem by utilizing all attributes as is, and directly incorporating them into the pattern mining process. In order to implement our method, we modify an existing graph-based knowledge discovery algorithm, SUBDUE, by adding it the capability of working with continuous and discrete data vectors of any dimension. In addition, we propose an object detection framework using improved SUBDUE in its object matching step. This system detects repetitive objects such as buildings and airplanes in satellite images, once the user specifies a sample target by drawing a bounding box around it. Experiments on artificial and real datasets show that our contributions result from a robust and flexible approach that can generalize over a vast number of problems.