As for low-resolution face detection in real-world video surveillance， achieving balance in terms of speed， accuracy， and memory consumption is of great importance due to the hardware constraints. Towards the problem， inspired by the more recent RetinaFace this paper proposes a light-weight scale-adaptive deep face detection model， termed as DLFace. Firstly， the improved depthwise separable convolution can effectively prevent information loss during training. Secondly， the improved deformable convolution is introduced into the backbone network and single stage headless （SSH） face detector， so as to enlarge the receptive field while also to adapt to facial changes such as expression， pose and so on. Finally， a Lambda layer is introduced in the high level of the backbone network， attempting to effectively explore the semantic and location information to form a richer representation of facial features. Experimental results on the WiderFace dataset show that DLFace has achieved a comparable or even better performance than existing light-weight face detection methods. Meanwhile， DLFace also achieves a better performance balance than most of previous methods in prediction efficiency and effectiveness.