Authentic Versus Counterfeit Image Classification after Re-Sampling and Compression

Abstract

Image-based applications such as inspection, quality assurance, print quality assessment, and authentication are often bandwidth limited, and benefit from image down-sampling or compression. Often the down-sampling or compression strategy is selected without considering the impact on the functional goal of the imaging task. In this paper, we use a specific "functional" imaging application–the distinction between images of authentic products and counterfeit products–to assess the impact of down-sampling and compression on the classification accuracy of the counterfeit detection imaging software. Our samples are originally scanned at high resolution (600 dots/in, or dpi) and then down-sampled to as little as 10 ppi (continuous tone pixels/in). JPEG compression to 2% and 1% of the original image size were performed separately or in combination with the various down-sampling factors. A simple Gaussian classifier was used throughout. We found that improved classification accuracy was obtained even for samples compressed or down-sampled by more than 99%. The implications of these findings for improved inspection results, authentication accuracy and counterfeit detection are then discussed.