Texton Theory Revisited. A computational colour approach
Place: Large Lecture Room, Computer Vision Center
Affiliation: Computer Vision Center/ Univ. Autònoma de Barcelona, Spain
Colour and texture are important visual cues for image understanding.
The definition of computational descriptors that combine both features is still an open problem. The difficulty is essentially due to the inherent nature of both cues, while texture is a property of region, colour is a property of a point.
Since now three approaches have been used for combining cues, (a) texture is a directly described in each one of the colour channels, (b) texture and colour are described separately and combined in a latter step, and (c) the combination is done using machine learning techniques.
Considering that this issue is solved at early stages of the human visual system, in this work we propose to study the problem using a direct implementation of a perceptual theory, the /texton/ theory, and to explore its extension to colour.
Since texton theory is based on the description of texture by the densities of local attributes, this matches perfectly with an holistic framework where descriptors are based on bag-of-words. Some descriptors based on different textons spaces and different image representations have been studied. Furthermore, the feasibility of these descriptors has also been studied for intermediate levels of image representation.
The proposed descriptors have proved high efficiency in retrieval and image classification. They also present some advantages in vocabulary generation. The quantification is done directly on low-dimensional spaces, whose perceptual properties allow low-level semantic associations to the visual words. The results make us to conclude that although the performance of holistic approaches is high, the introduction of spatial co-ocurrence of blob properties, shape and colour, is a key element for their combination. This conclusion agrees with perceptual evidences.