Automatic Annotation of Images by a Statistical Learning Approach

A novel statistical learning approach for automatic annotation of images is presented. A minimum probability of error annotation is feasible with our approach. Firstly, an image is represented as a bag of feature vectors by dividing the image into small blocks, from each of which a six-dimension feature vector is extracted. Secondly, we established the probabilistic formulation for automatic annotation of images through estimating the Gaussian mixtures of each image and the common Gaussian mixtures of all the images with the same semantic label. At last, the steps of the training and annotation algorithm are given based on our probabilistic formulation. Experimental results show that the proposed supervised formulation achieve higher accuracy than previously published method.