Adaptive visually lossless JPEG-based color image compression

The paper presents two approaches to adaptive JPEG-based compression of color images inside digital cameras. Compression for both approaches, although lossy, is organized in such a manner that introduced distortions are not visible. This is done taking into account quality of each original image before it is subject to lossy compression. Noise characteristics and blur are assumed to be the main factors determining visual quality of original images. They are estimated in a fast and blind (automatic) manner for images in RAW format (first approach) and in Bitmap (second approach). The dominant distorting factor which can be either noise or blur is determined. Then, the scaling factor (SF) of JPEG quantization table is adaptively adjusted to preserve valuable information in a compressed image with taking into account estimated noise and blur influence. The advantages and drawbacks of the proposed approaches are discussed. Both approaches are intensively tested for real-life images. It is demonstrated that the second approach provides more accurate estimate of degrading factor characteristics, and thus, a larger compression ratio (CR) increase compared to super-high quality (SHQ) mode used in consumer digital cameras. The first approach mainly relies on the prediction of noise and blur characteristics to be observed in Bitmap images after a set of nonlinear operations applied to RAW data in image processing chain. It is simpler and requires less memory but appeared to be slightly less beneficial. Both approaches are shown to provide, on the average, more than two times increase in average CR compared to SHQ mode without introducing visible distortions with respect to SHQ compressed images. This is proven by the analysis of modern visual quality metrics able to adequately characterize compressed image quality.