Font Recognition Using Shape-Based Quad-tree and Kd-tree Decomposition

The search for appropriate data representations and visual features for content-based image retrieval continues within the computer vision community, alongside the development of new matching and indexing techniques to facilitate fast search in large-scale image databases. In this study, we present a solution to the problem of typeface identification and character recognition in text-based images using this type of approach. Geometrical properties of a character are extracted from its binary image at different levels of spatial resolution, via a hierarchical abstraction of the image data. Two such abstractions are described here: a shape-based quad-tree, and a kd-tree. Unlike the traditional quad-tree representation in which an image is generally partitioned into 4 blocks of equal size at each level of decomposition, the block size in the centroidbased quad-tree is variable, being determined by the location of the centre of gravity of the regions represented in a sub-image. In a similar way, the kd-tree partitions the image data into two, again about an axis defined by the shape of the region represented in the image. Weighted and non-weighted feature vectors of the partitioning points are then used within a metric tree to index character images in a font database. We discuss factors that influence the performance of the resulting font retrieval system, both in terms of accuracy and speed.