Vector quantization (VQ) techniques for image coding are sensitive to training data. Significant differences between training data and test data result in poor image representation. A simple example of this is a VQ codebook trained on smooth, untextured images. When used to encode an image with hard edges, the most similar codebook entry offers poor representation of an edge. This task tests this training data sensitivity when VQ is used, not in isolation, but as a third stage of an image compression system. The first two stages being Wavelet decomposition and texture classification.
A simulation of this algorithm has been implemented. Since a great number of training and test images are needed, a survey of Internet sites was conducted. A collection of 669 medical images were collected off the Internet and classified by subject matter in to 49 categories. The first experiment, classifying images by imaging mode and body part, has been completed.
Other schemes for classification are possible and will be explored.