Abstract: Aimed at the low accuracy of occluded face detection in complex scenes, an occlusion face detection method combining depth dictionary learning and feature reconstruction is proposed. A shallow CNN was used to generate face candidate regions, and the pre-trained VGG16 network was used to characterize them. A sparse coding method was used to establish a deep retrieval dictionary composed of typical faces and non-faces. Using the locality preserving projections method, the feature descriptor of the face candidate region was reconstructed into a similarity-based feature vector by using the retrieval dictionary. The reconstructed feature vector was sent to the deep neural network to perform face/nonface classification and face bounding box location regression simultaneously. The experimental results on the MAFA occlusion face dataset show that the detection accuracy of this method is about 12.3 percentage points higher than the current mainstream face detection method.