Custom Lossless Compression and High-Quality Lossy Compression of White Blood Cell Microscopy Images for Display and Machine Learning Applications

Abstract: This master's thesis investigates both custom lossless compression and high-quality lossy compression of microscopy images of white blood cells produced by CellaVision's blood analysis systems. A number of different compression strategies have been developed and evaluated, all of which are taking advantage of the specific color filter array used in the sensor in the cameras in the analysis systems. Lossless compression has been the main focus of this thesis. The lossless compression method, of those developed, that gave best result is based on a statistical autoregressive model. A model is constructed for each color channel with external information from the other color channels. The difference between the predictions from the statistical model and the original is further Huffman coded. The method achieves an average bit-rate of 3.0409 bits per pixel on the test set consisting of 604 images. The proposed lossy method is based on taking the difference between the image compressed with an ordinary lossy compression method, JPEG 2000, and the original image. The JPEG 2000 image is saved, as well as the differences at the foreground (i.e. locations with cells), in order to keep the cells identical to the cells in the original image, but allow loss of information for the, not so important, background. This method achieves a bit-rate of 2.4451 bits per pixel, with a peak signal-to-noise-ratio (PSNR) of 48.05 dB.