Efficient Face Detection Under Complex Conditions

With the development of computer vision technologies,more and more face related applications are applied in real life.Face detection is the basis of all other face related tasks.In recent years,face detection technology based on convolution neural networks has made great progress.However,due to the diversities of illumination,distance and angle,face detection in unconstrained environment is still a challenge.We propose specific solutions to two difficulties in face detection.To solve the problem that faces are hard to be detected in unconstrained environment,we propose a highly accurate face detector based on feature fusion.By fusing the features in adjacent levels in the feature pyramid,our feature fusion module can make use of the potential correlations among hierarchy features.Spatial attention mechanism is also introduced to boost the performance of face detection.The experimental results show that the proposed model achieves a significant increase in AP score without additional computational burden.To address the inefficiency of face detectors,we propose an edge device face detector based on binarized convolution neural networks.Through the binary quantization of activations and weights in a network,we reduce the space and computation cost by a large margin while keeping acceptable accuracy.