Image annotation is a task of assigning a set of semantic tags or keywords to unlabeled image based on a training set of data. This machine learning process depends on extraction and clustering the lowlevel features of images then mapping them to the semantic which is high-interest image retrieval. Many annotation techniques with reasonable performance have proposed in the last decade. The proposed algorithm for automatic annotation in this paper depends on similarity computation and label transferring to the query image. Similarity computing uses low-level image features and significant distance estimates. The feature vector of the test image compared with the feature matrix of similar and dissimilar image pairs in the training data sets. Then the label or keyword transfer from the similar image pair to the test image is performed by counting the local frequency of neighbor's keywords. Performance is evaluated using precision and recall.