Modified JPEG scheme to compress reconnaissance images

The JPEG standard is one of the most popular image compression schemes today. However, JPEG usually compresses the whole image equally, while in our application of air-to-ground reconnaissance images and many others, some parts of the image are much more important than the rest. Hence, it is logical to compress different parts of the image differently by trading the quality of background for improved quality of objects of interest. We develop an algorithm based on the JPEG baseline to achieve this goal. Since JPEG already uses a block-based discrete cosine transform (BDCT), we design a block classifier based on the BDCT coefficient values of the block being processed and its neighboring blocks. This approach is feasible, as our objects of interest are fabricated structures, roads, and vehicles possessing certain sizes and distinctive boundaries from the background, which follow certain patterns in the BDCT domain. We also use a projection onto convex sets (POCS) algorithm in the decoder to remove the artificial blocking effects. Therefore, we adjust the POCS procedure with respect to the classifier as well. Experimental results are given to compare the performances of the proposed scheme with the conventional JPEG scheme.