Beyond "Near Duplicates": Learning Hash Codes for Efficient Similar-Image Retrieval