Using Colour Gabor Texture Features for Scene Understanding

Gabor lters have been used extensively as a model of texture for image interpretation tasks. This paper demonstrates that when a bank of Gabor lters is applied to an image, there are strong relationships between the outputs of the di erent lters. These relationships are used to devise a new texture feature which is capable of describing texture information in a concise manner. Information about the distributions of lter responses is also encoded in the new feature. Performance of the feature is assessed by applying it to an image region classi cation task and comparing results to those obtained using features which do not utilise the relationships between lter outputs. It is shown that the distribution information aids the classi cation task. The new feature performs comparably with the other features whilst yielding a signi cantly smaller feature vector. We then describe how the feature may be applied to colour images. It is shown that the inclusion of colour information is bene cial to the classi cation task and also that the choice of colour space is important. The classi cation results are then compared to those obtained using a 28 element feature encoding colour, position, shape, size, context and also texture. The new colour Gabor feature outperforms the more intuitive 28 element feature. We conclude by suggesting that the Gabor based feature may be capable of implicitly encoding some shape and context information.