Face Detection Algorithm Based On Context Information In Unconstrained Scenes

Face detection has important research value in the fields of identity authentication,video surveillance,attendance system,emotion analysis,etc.It is also an important research topic in computer vision.In the constraint scenario,the existing face detection method has achieved a high detection speed and accuracy;however,the face in the real scene is usually affected by many factors,such as a large range of size changes,the different lighting conditions and the problem of facial occlusion,which brings many difficulties and challenges to the realization of more accurate and efficient face detection methods.So,it is imperative to design a face detection algorithm that can cope with the interference caused by complex scenes.In this paper,an efficient multi-scale face detection algorithm is designed for the difficulties in face detection:Firstly,we use a single-stage network model from the perspective of structural design,the features of different semantic levels are used to realize the targeted detection of faces at different scales.Especially for small-scale face branches,the weights sharing and feature fusion structure are used to enhance the feature expression of candidate targets and improve the ability of network to detect targets.Secondly,the skip connection is used to implement a context sensitive module with multiple receptive fields and multiple semantic levels.It enhances the feature transfer between different semantic levels,enriches the feature representation of candidate targets,and utilizes cavity convolution in the backbone network to extract global information of candidate targets to help the network achieve better classification and regression.Finally,for the problem of class imbalance in small-scale target detection,the Focal Loss and ratio of scale specialization is used in the training phase,which use the difficulty level of the samples to control the loss weight in the optimization process.While focusing on the difficult samples for network optimization,it also retains the optimized contribution of easy samples,and improves the discriminating ability of the detection network for difficult samples.