Research And Implementation Of Face Detection Based On Incomplete Information

Face detection refers to the process of determining whether there are faces in the input image and the specific locations where they exist.With the development of intelligence,face detection technology is widely used in the intelligent interaction,identity verification,mobile social networks and other fields.However,faces in input images are not always clear and complete under unrestricted conditions.Face images with incomplete information caused by occlusion,angle,resolution and other factors pose great challenges on face detection.Although many algorithms have achieved good results in face detection,there are still some shortcomings in dealing with incomplete information faces.In this thesis,an improved algorithm is proposed to improve the detection accuracy for the incomplete face caused by occlusion and small scale.The main contributions of this thesis are as follows:(1)A feature-enhanced subnet SG-net based on region generation is proposed.The lack of partial facial features caused by occlusion reduces the accuracy of current face classification techniques.Inspired by the idea of attention mechanism,this thesis proposes a feature enhancement subnet based on target region generation,which applies features in visible face regions for face classification.By constructing the target data set,the SG-subnet is trained to recover the region near the face from the original image features.Then an improved SG-Faster RCNN model is proposed by fusing the generated region features with the original image convolution features.Experiments show that the accuracy of SG-Faster RCNN combined with SG-net is 4.4% higher than that of Faster RCNN.Feature fusion based on SG-net can effectively improve the detection effect of face with incomplete information caused by occlusion.(2)A multi-scale face detection model FSG-FD based on feature enhancement is proposed.At present,the detection models based on single-scale features have the problem of inaccuracy in small face detection.In order to improve the accuracy of occlusion and small face detection,this thesis proposes a FSG-FD detection model.By introducing multi-scale feature pyramid structure,fusing high-level features of high semantic information and low-level features of high resolution,and combining feature enhancement branch SG-net,the accuracy of the model is improved by 4.3% compared with SG-Faster RCNN on Wider dataset.In addition,this paper simulates practical application scenarios by manually annotated monitoring data sets.The detection accuracy of FSG-FD on monitoring data sets is 85.1%.(3)A face detection system for security monitoring has been designed and implemented.Security monitoring is an important application field of face detection technology.Faces captured in natural environment often include incomplete faces caused by various occlusion,angle and other factors.Aiming at the application requirement of security surveillance scenarios,this thesis designs and implements a face detection system which supports video and image based on the FSG-FD detection model proposed in this paper.