Résumé : This paper considers reference free quality assessment of distorted and noisy images. Specifically, it considers the first and second order statistics of stem noise that can be evaluated given any image. In the research field of Image quality Assessment (IQA), the stem noise is defined as the input of an Auto Regressive (AR) process, from which a low-energy and de-correlated version of the image can be recovered. To estimate the AR model parameters and associated stem noise energy, the Yule-walker equations are used such that the accompanying Auto Correlation Function (ACF) coefficients can be treated as model parameters for image reconstruction. To characterize systematic signal dependent and signal independent distortions, the mean and variance of stem noise can be evaluated over the image. Crucially, this paper shows that these statistics have a predictive validity in relation to human ratings of image quality. Furthermore, under certain kinds of image distortion, stem noise statistics show very significant correlations with established measures of image quality.