Evaluating human brain potentials during watching different images can be used for memory evaluation, information retrieving, guilty-innocent identification and examining the brain response. In this study, the effects of watching images, with different levels of familiarity, on subjects' Electroencephalogram (EEG) have been studied. Three different groups of images with three familiarity levels of "unfamiliar", "familiar" and "very familiar" have been considered for this study. EEG signals of 21 subjects (14 men) were recorded. After signal acquisition, pre-processing, including noise and artifact removal, were performed on epochs of data. Features, including spatial-statistical, wavelet, frequency and harmonic parameters, and also correlation between recording channels, were extracted from the data. Then, we evaluated the efficiency of the extracted features by using p-value and also an orthogonal feature selection method (combination of Gram-Schmitt method and Fisher discriminant ratio) for feature dimensional reduction. As the final step of feature selection, we used 'add-r take-away l' method for choosing the most discriminative features. For data classification, including all two-class and three-class cases, we applied Support Vector Machine (SVM) on the extracted features. The correct classification rates (CCR) for "unfamiliar-familiar", "unfamiliar-very familiar" and "familiar-very familiar" cases were 85.6%, 92.6%, and 70.6%, respectively. The best results of classifications were obtained in pre-frontal and frontal regions of brain. Also, wavelet, frequency and harmonic features were among the most discriminative features. Finally, in three-class case, the best CCR was 86.8%.