Research On Multi-scale Face Detection Based On Convolution Neural Networks

Face detection is a key technology in face information processing,which has become a research field in computer vision.In recent years,face detection has been received widespread attention and widely used in different fields.Aiming at the low average precision of multi-scale face detection in unconstrained environment,we analyze Single Stage Headless face detector,and propose its optimization scheme.The new method improves the performance of multi-scale face detection.Firstly,we introduce six typical face detection models and describes the network architecture in detail.In order to further improve face detection performance,we select Single Stage Headless face detector as the basic model to alleviate the problem of low average precision for small-scale face detection.Secondly,we propose a multi-scale face detection method.Considering the importance of context information for face localization and regression,the recurrent neural network modules initialized with the unit matrix are added after the last convolutional layer of VGG-16.On the basis of original multi-layer fusion,we adopt more multi-layer fusion schemes based on assigning weights to extract the context information of face.The scheme improves the performance of face detection.Thirdly,during inference,the predicted boxed from different scales are joined together followed by Soft Non-Maximum Suppression to form the final detections.The algorithm with the best function and parameters can alleviate the occlusion problem of face detection.Finally,in order to verify the effectiveness of the proposed method,a lot of experiments are performed on WIDER FACE dataset.Our method outperforms SSH on “easy”,“medium” and “hard” subsets of WIDER FACE dataset and improves the average precision by “0.2%”,“0.5%” and “2.0%” on the validation set.At the same time,the illumination and occlusion experiments were performed on the SoF dataset.The results show that the improved model is more accurate than the original SSH for face locating.The proposed method can effectively improve the performance of multi-scale face detection in unconstrained environment by adding recurrent neural network modules initialized with the unit matrix,applying more multi-layer fusion scheme and Soft Non-Maximum Suppression algorithm.