The Vector Quantisation Segmentation Method
In our work, we employ the texture based segmentation algorithm to define a complete partition of an image that leads to modelling of the textures as polynomial surfaces. In order to achieve the segmentation and partitioning of the image, a Vector Quantisation (VQ) segmentation method is considered which is based on the use of appropriate training data. The basic idea of this method is summarised in the following paragraphs.
A feature vector is assigned to every point containing statistical information of that point. This information can represent the mean, standard deviation and contrast measured both on image data and on differences of image data within a window of 7x7. The statistical information is contained in the transform coefficients derived from DCT. In general, a quantiser reduces the number of bits required to store the transformed coefficients by reducing the precision of those values. This storage can be perceived by the introduction of a mapping q that reproduces an input vector
by a reproduction vector
(
) drawn from a finite reproduction alphabet called codebook. An efficient quantiser is said to be one which is optimal, or equivalently one for which the distortion caused by the reproduction of the input feature vectors is minimum. Certainly, the design of an efficient quantiser is an optimisation problem and it is usually necessary to know the probability density function of the input vectors.
The LGB algorithm is subsequently employed in order to perform the vector quantisation. This algorithm differs from previous approaches in that it is not based on variational and differentiational techniques and it can be used both on a known probabilistic model or on a long training sequence of data.
According to the LGB algorithm, a training set of feature vectors are provided (input feature vectors) and a selection of an initial set of prototype (codewords) vectors is made. The distance between the training sequence feature vectors and the initial prototype vectors determines whether the latter will be accepted to be the prototype vectors of the quantiser. In addition, it helps to allocate each training vector to its closest prototype vector. If the distance which is associated with the distortion measure is smaller than a threshold, the prototype vectors introduced are accepted otherwise, we replace a prototype vector
with a vector situated at the centroid of the subset of the training vectors that were mapped to into
during the previous stage. Then, the same procedure discussed in this paragraph is followed until the distortion is smaller than the threshold.
By means of the above method, we succeed in creating a set of vectors that can be used to assign a label or a category to every feature point in the image. This produces a labelled image from which a set of polygons is extracted.
The code built is available both in a html and a C++ format.
References
Last updated 19 November 2000