To address the local occlusion and pose variation in face detection, face can be looked on as a whole composed of several parts from up to down. First, the face is divided into a number of local regions from which various features are extracted. Each region is identified by a local classifier and is assigned a preliminary part label. A random field is established based on these labels and multiple dependencies between different parts are modeled in a CRF framework. The probability that the test image may be a face is calculated by a trained CRF model. The probability is used as a measure to test the existence of a face. The experiments were carried out on the CMU/MIT dataset. As indicated by the experiment results, the following methods can improve the detection rate and enhance the robustness of face detection in case of occlusion: 1) integrating multiple features and multiple dependencies in CRF framework; 2) dividing the face optimally.