In this paper we consider the problem of automatically annotating images with keywords. We first discuss performance measures for the problem in some length. We propose a new information-theory based measure de-symmetrised mutual information (DTMI). We then describe a straightforward solution to the annotation problem. We first train a set of classifiers to detect the presence of each individual keyword in the set of training images. For this we use the PicSOM image analysis framework. We then describe a method of converting the classifier outputs back into keyword annotations for the test set. We compare the performance of the proposed method experimentally to that of other methods presented in the literature. For the experiments we use data from the Corel database. The result of the comparison is favourable to the proposed method.