Matched polynomial features for the analysis of grayscale biomedical images

We considered the general form of polynomial features represented as polynomials in the image pixels domain. We showed that under natural constraints these polynomial features turned to linear combinations of the image autocovariance function readings. We proposed a number of approaches for matching the features under study with texture properties of images from a training sample. During computational experiments on three sets of real diagnostic images we demonstrated the efficiency of the proposed features, which involved the decrease of the recognition error probability of X-ray bone tissue images from 0.10 down to 0.06 in comparison with the previously studied methods.