Face detection is a hot topic in computer vision. The face detectionmethods usually consist of two subtasks, i.e. the classification sub-task and the regression subtask, which are trained with differentsamples. However, current face detection knowledge distillationmethods usually couple the two subtasks, and use the same set ofsamples in the distillation task. In this paper, we propose a taskdecoupled knowledge distillation method, which decouples the de-tection distillation task into two subtasks and uses different samplesin distilling the features of different subtasks. We firstly propose afeature decoupling method to decouple the classification featuresand the regression features, without introducing any extra calcula-tions at inference time. Specifically, we generate the correspondingfeatures by adding task-specific convolutions in the teacher net-work and adding adaption convolutions on the feature maps ofthe student network. Then we select different samples for differentsubtasks to imitate. Moreover, we also propose an effective proba-bility distillation method to joint boost the accuracy of the studentnetwork. We apply our distillation method on a lightweight face de-tector, EagleEye[37]. Experimental results show that the proposedmethod effectively improves the student detector’s accuracy by5.1%, 5.1%, and 2.8% AP in Easy, Medium, Hard subsets respectively.