In this paper we propose a new prediction scheme for lossless compression of fingerprint images. The proposed method uses the symmetric/near symmetric property of such images. The scheme first divides the image horizontally into two halves and then the upper half is divided vertically in another two halves; resulting in three parts of the given image. It is observed that the two halves of the upper part are highly correlated. In view of this observation, the pixels of the right half are predicted using those in the left half and also in the present half. On the other hand, pixels of the left half and those of the bottom half are classified as per slope bins of GAP and then LS based predictor is used with respective classes. Prediction performance of the proposed scheme is found to be significantly better than many competitive and popularly used methods. © 2012 IEEE.