As a cooperation project between Sweden and China, we sampled a number of rock specimens for analyze rock fracture network by optical image technique. The samples are resin injected, in which way; opened fractures can be seen clearly by means of UV (Ultraviolet) light illumination. In the study period, Recognition of rock fractures is crucial in many rock engineering applications. In order to successfully applying automatic image processing techniques for the problem of automatic (or semi-automatic) rock fracture detection and description, the key (and hardest task) is the automatic detection of fractures robustly in images. When statistical pattern recognition is used to segment a rock joint color image, features of different samples can be learned first, then, each pixel of the image is classified by these features. As the testing result showing, an attribute rock fracture image is segmented satisfactorily by using this way. The method can be widely used for other complicated images too. In this paper, Kernel Fisher discrimination (KFD) is employed to construct a statistical pattern recognition classifier. KFD can transform nonlinear discrimination in an attribute space with high dimension, into linear discrimination in a feature space with low dimension. While one needs not know the detailed mapping form from attribute space to feature space in the process of transformation. It is proved that this method performs well by segmenting complicated rock joint color images.