Segment Feature Co-ocurrence Based Texture Detection

This work on texture detection was inspired by the general problem of object recognition in two dimensional still images. One of the crucial challenges associated with the object recognition is selecting and obtaining discriminative features. In analysis of real scenes, like nature or urban places, many objects contain textures, which can be considered as one of the object features. Texture detection may also significantly improve image segmentation, which is one of the tools used for object recognition. This report presents a novel method named Feature Co-ocurence Texture Detector (FCTD) which allows for fully automatic detection of textures common in real scenes, like water in lakes, tiger skin, fields of flowers or tree crowns. The method searches for an alternating color pattern, like for example black and orange stripes covering tiger skin, which is very often present in those textures. The final result is a hierarchy of textures (described by their boundaries and a set of features) detected at multiple precision levels, which can be used for further analysis or texture classification. It is achieved through hierarchical clustering of color pairs related to adjacent image segments, where each segment represents a low color gradient, simple shaped patch of pixels in the image. The results are presented on some images from the Berkeley database.