Research Of Face Detection And Face Alignment Algorithms In Uncontrolled Conditions

At present,face detection and face alignment are basically practical in situations such as good lighting,unobstructed face,no extreme face pose,and high image resolution.For example,comparison of witnesses,beauty cameras,etc.Face detection alignment application at close range.However,face alignment and face detection in uncontrolled scenes are still not satisfactory,especially in the face of complex facial expressions,serious face occlusion,complex external environment background,poor lighting conditions,low image resolution,low face resolution,and so on.Face alignment and face detection cannot be achieved better,and real-time detection is more difficult.With the increasing maturity of artificial intelligence algorithms such as deep learning,the face alignment and face detection technologies in uncontrolled scenes have been greatly improved.This topic is based on the convolutional neural network to study the face alignment algorithm and bottom-up face detection algorithm in uncontrolled scenarios.The main work is as follows:1.The three face alignment algorithms based on model matching,cascade-based regression,deep learning are compared and analyzed,and a face alignment algorithm based on convolutional neural network is proposed.Using multi-task learning strategies and training data augmentation strategies,using dense networks and six cascaded networks for face-related key points detection and key points location relationship detection,face alignment algorithms based on convolutional neural networks are implemented.Experimental results show that under different postures,different degrees of occlusion,and different lighting conditions,the subject algorithm has more obvious advantages than other algorithms in uncontrolled conditions,and the average error in the published data set(AFLW)is 6.8%.2.The three face detection algorithms based on cascading,deformable component model and deep learning are compared and analyzed.A bottom-up face detection algorithm is proposed.The face-related key points coordinates are firstly used to calculate the face frame coordinates based on the knowledge method,and a bottom-up face detection algorithm is implemented.Experimental results show that the proposed algorithm is more robust than other algorithms under different postures,different occlusion levels,and different lighting conditions,and the detection time is less affected by the number of faces in a picture.According to the discontinuous receiver operating characteristic curve from the published data set(FDDB),there are 1,500 false detection faces with a recall of 0.98.3.In order to achieve real-time face detection and face alignment under uncontrolled conditions,multi-thread parallel processing is used,and the number of cascaded networks is optimized.Experimental results show that the discontinuous receiver operating characteristic curve from the published data set(FDDB),there are 1,000 false detection faces,the recall rate of using two cascaded networks is lower by 0.1 than that of the six cascaded networks,detection speed has greatly improved.Through actual tests in the monitoring scenario,the input image resolution is 1920?1080,using an image processor(Nvidia Ge Fore GTX 1070),CPU is Inter Core i7-7700HQ@2.8GHz,4 cores and 8 threads,and the minimum face resolution can be detected as 12 ?12.The average detection frame rate is 10.2fps.Compared with other algorithms,this thesis algorithm has a certain practical value.