Research On Face Detection In Complex Background

Complex background is a background with interference factors such as occlusion,illumination,blurring and different facial postures.The interference factors in complex background can lead to inaccurate features of human faces,which makes the research of face detection in complex background very difficult.At present,more and more scenes need face detection in complex background,such as driverless cars and national anti-terrorism.These requirements make the research of face detection in complex background have important practical significance.At the same time,in theoretical research,face detection under complex background is the key research to simulate human vision,and is an important part of the realization of artificial intelligence for robots.Therefore,this research has important theoretical significance.In this paper,we mainly focus on improving the accuracy and speed of face detection in complex background:1)The general target detection model Faster R-CNN is explained in detail.Faster R-CNN model has excellent detection effect in the field of general target detection.Especially the RPN network proposed by this model can generate high-quality candidate regions,and then obtain high-quality feature vectors to improve the detection effect of the model.At the same time,the model has a good experimental effect in detecting small objects,and the face in complex background is mostly small faces.Therefore,this paper uses Faster R-CNN model as the basic model to improve.2)In this paper,a face detection model based on environment information and feature fusion is proposed.The model uses residual network as feature extractor to extract features for classification and location regression.Meanwhile,feature fusion is used to make features for classification and regression have more location information and improve the positioning effect of the model.Finally,the model improves the environmental information proposed in the CMS-RCNN model,considering that the face in complex background is usually atypical posture,the proportion of face to body will be reduced accordingly.In this paper,the ratio of 0.75 times in CMS-RCNN model is set by experiment.The experimental results show that the model achieves high detection accuracy.Meanwhile,the average detection speed of the CMS-RCNN model is 1 FPS.The average detection speed of the proposed model is 4 FPS.3)In this paper,we presents a three-class face detection model based on Faster R-CNN model.Faster R-CNN model is used as the basic model.Faces are clearly divided into small faces and normal faces,so that the model pays more attention to small faces.Then the image pyramid is used to enhance model for small face detection.At the same time,the feature fusion mechanism designed in HR model is used to improve the effect of feature extraction.The experimental results show that compared with the most advanced HR model,the proposed model is very close to the HR model in terms of detection accuracy,and in terms of detection speed,the average detection speed of the HR model is 0.5 FPS,while the average detection speed of the model reaches 6 FPS.