Image compression

Since epitome modeling gives a generative model for the image and the epitome is significantly smaller than the original image, using epitome in image compression seems to be promising. The basic approach is:

• At the encoder side, perform epitome modeling. Compress the epitome mean image using some existing lossy compression algorithm, and compress the mapping information using lossless techniques, such as variable length coding or arithmetic coding.
• In the decoder size, perform reconstruction as decompression.

We use JPEG to compress the epitome mean image, and left the mapping information compressed. So the number of bits = bits used in compressing epitome mean + number of patches * log2(number of possible mappings). We compare this with the JPEG algorithm applying on the original image, and the performance is shown as rate-distortion curves in terms of bitrates vs. PSNR (peak signal to noise ratio) as the following figure:

 The performance of epitome-based compression is much worse than JPEG (take more bits for the same quality), which is not surprising since patches have many overlaps and the mapping information is uncompressed. One thing we found in simulation is: epitome-based reconstruction have better visual quality than JPEG-based algorithm for the same PSNR value, here is an example:

         

                 (a)                                   (b)                                    (c)

(a) original image (b) jpeg compressed image with 22.77 dB (c) epitome-compressed with 21.93 dB

There are some practical issues in such a system:

• How large is the epitome size?
• How to pick patches from image and how many patches to take?
• A good rate-distortion framework is essential for a successful coding system. I.e., for a given bitrate, how to optimize the decoded image quality. It’s not entirely clear how to come up with such a framework for the mapping information.

And there are some obvious disadvantages in such an epitome-based image compression system:

• The epitome works (in terms of reconstruction quality) on images with a lot of textures. For heavily structured images or images with large texture variations, the reconstruction quality is poor, and that poses an upper bound for the epitome-based compression scheme.
• The EM-based learning algorithm is very computationally expensive, which practically makes it uncompetitive with existing image coders.

Using epitome in image compression is only a crude and tentative idea, and we believe there are a lot of room to improve.