With wide spread availability of digital cameras, smart phones and other imaging devices, large amount of similar images are routinely created. Such images need to be compressed for storage and communication purposes. Instead of compressing each image individually, it is possible to achieve higher compression efficiency by compressing the whole set of similar images together. Any such similar image compression scheme needs to support random access as users may want to access the images in any random order. To reduce the inter-image redundancy, one key problem is the construction of predictors to minimize the inter-image redundancy. The existing methods are not effective in their predictors and intra prediction design....[ Read more ]

With wide spread availability of digital cameras, smart phones and other imaging devices, large amount of similar images are routinely created. Such images need to be compressed for storage and communication purposes. Instead of compressing each image individually, it is possible to achieve higher compression efficiency by compressing the whole set of similar images together. Any such similar image compression scheme needs to support random access as users may want to access the images in any random order. To reduce the inter-image redundancy, one key problem is the construction of predictors to minimize the inter-image redundancy. The existing methods are not effective in their predictors and intra prediction design.

In this thesis, we propose a compression method for medical images based on low frequency features of the images. We observe that the pairwise similarities among the similar images are not uniform among the images. Some pairs have similarity for a relatively small frequency band only. Some have similarity for wider frequency bands. Motivated by this, we propose a set of multi-level low frequency templates as predictors for each image. The resulting difference images tend to have lower entropy and require fewer bits to encode than the single-level low frequency template method. Results suggest that up to 30% of gain is possible. The performance gain is more obvious in larger sets of similar images with higher dissimilarity.

We also propose an effective compression scheme for photo albums based on global/local motion estimation and compensation, and intra prediction. We observe that many photos in photo albums are very similar in the sense that basically they capture the same scene with different camera zoom factors, different rotational positions and different translational positions. The global motion estimation and compensation can account for such camera motion among similar photos. On the other hand, the local motion estimation and compensation can account for object motions. Experiments suggest that this algorithm can achieve considerable performance gain.

Index Terms— Image compression, set redundancy compression, photo album compression