Metrics for evaluation of lossy compression effects on digit recognition in SMD images

In this paper we define metrics for objective evaluation of the degradation caused by lossy compression on a specific image processing task. The study is done on gray-scale Surface Mounted Device (SMD) images degraded by lossy JPEG compression. We focus on evaluating the impact of the distortion on the extraction of a set of digits which form a label on top of the SMD component. A digit extraction systems is devised in order to isolate them and produce individual binary masks. Our approach is to compare the results of a simple digit extraction system from the uncompressed (reference) image with the results obtained from the compressed/decompressed version. The use our previously devised binary metrics to evaluate the distortion of the resulting binary masks of the digits. We then define and apply a new metric that directly measures the distortion as a distance in a feature space. The digit extraction is then improved using region growing and the compression distortion is evaluated using the new feature metric. The results show that this metric does not change considerably over a large range of compression rates. This metric also provides objective way of evaluating performance of bath processing methods in the presence of a compression/decompression step