JPEG-deblocking of Blood Cell Images using Deep Learning

Abstract: This thesis investigates the use of convolutional neural networks as a reconstruction or JPEG- deblocking model for JPEG-compressed blood cell images, needed due to the well known block artifacts caused by JPEG-compression. CellaVision develops automated microscopy for blood analysis that detects and classifies blood cells from images. The automated analysis is carried out on high resolu- tion microscope images, before the images are compressed to JPEG-75 format, where 75 is the quality factor. We investigate how hard the blood cell images can be compressed to still enable acceptable reconstruction quality for display to the user. We propose a CNN-model that reconstructs blood cell JPEG-images of quality factor 50 and higher, to PSNR and SSIM values on average higher than JPEG-75 blood cell images. Using our method, 99.9% of the blood cell images are improved in terms of PSNR and SSIM. However, these metrics do not take into account opinions of professional laboratory technicians who are the main users of CellaVision’s application. A comparison is made between a model predicting in the RGB colorspace and the YCbCr colorspace, the later being exploited by the JPEG-compression algorithm. Results show that models trained on input images of higher or random quality outperform models trained on lower quality, even in the reconstruction of low quality images. Predicting from a higher quality factor is safer when considering quality criteria for medical images and their use in diagnosing, where image quality is critical. With our thesis we propose that CellaVision could store the images with JPEG-50 and still achieve a reconstructed image with a quality as good as JPEG-75, based on the SSIM and PSNR metrics, and thereby save 31% of storage space compared to today.