Blind-estimating Algorithm Of Compressed Image Quality

Common image/video codec such as JPEG, MPEG-2 and MPEG-4 are all based on the Discrete Cosine Transform: Original image is first transformed into the DCT domain, where the coefficients will be quantized. As result, image can be described using fewer data bits, i.e., image is compressed. But this kind of compression will create side-effect: Image will suffer quality loss as the DCT coefficients been quantized and high-frequent part became coarse.PSNR (Peak Signal Noise Ratio) is a common objective dimension for the quality assessment of compressed or distorted images. The quality loss through compression can be estimated using PSNR. But traditional method, which calculates PSNR in spatial domain, needs the original uncompressed image as reference. This is the mainly restrict of the traditional method while original image is hardly available at the decode side. As the statistic of image coefficients in DCT domain indicates: the probability density function of DCT coefficients (AC term) can be approximated using a Laplace distribution with parameterλ. Using suchλ,and the quantization steps used, the difference (energy) between DCT coefficients of compressed and uncompressed images can be estimated. Further, with help of Parseval's theorem, the difference in spatial domain, and PSNR between 2 images, can be calculated.This thesis proposes a new method forλestimation, based on training, matching and image contents. Laplace parameterλcan be estimated without original uncompressed image. So will be the PSNR between compressed image and original one.Experimental results show that proposed method fits the JPEG image and frame from MPEG sequence well. When comparing with the existed algorithm using similar routine, proposed method also achieves a better evaluating result.