In recent years, Conditional Random Fields (CRF) are proposed and proved greatly useful in natural language processing, voice recognition and computer vision. In this paper we propose a variant of CRF to solve the problem of image binarization. Unlike previous image binariztion approaches, the Patch Random Fields (PRF) proposed here could provide global optimal solutions considering both the local information from source images and pixel-wise smoothness. In this new framework, we take image patch as a kind of raw information carrier and model it with mixture of probabilistic PCA. Moreover, traditional CRF always confronts difficulties in obtaining proper parameters for the probabilistic models; this process is often time-consuming and intractable. To mitigate this problem, we train most parameters in a generative way, and then optimize the remaining parameters using a gradient descent method. The advantages of generative models and CRF are thus well combined. Experimental results demonstrate our method's effectiveness.