It is an indispensable step to faithfully evaluate and control the perceptual quality of digital visual applications such as image compression, image restoration and multimedia streaming. In this paper, we propose an efficient and effective twostep framework named First No-reference Then Full-reference (FNTF), to evaluate different kinds of noise, and distinguish the quality of distorted image, using features Natural Scene Statistics of Dark Channel(NSSDC), Average Distance of Matched Keypoints(ADMK) and Dark Channel Similarity Deviation(DCSD) we proposed. Dark Channel is a kind of natural statistics based on the key observation - most local patches in images contain some pixels whose intensity are close to zero in at least one color channel. Features extracted from Dark Channel can greatly represent the pollution level of the image and the kind of noise it sufferd from. The average distance and dark channel similarity are sensitive to image distortions, while different local structures in a distorted image suffer different distance and degrees of similarity, respectively. This motivated us to explore global variation based local quality for overall image quality prediction. We find that our two-step framework can predict accurately perceptual image quality.