Research On Implementation And Application Of Fast Face Detection Algorithm Based On Deep Learning

Facial features,as an important biological identity,are a reliable way to achieve identity recognition.Traditional face recognition algorithms only focus on front faces,so the demand for face detection is not significant.With the development of face application scene,face detection has become an important field,especially in video content detection,real-time monitoring and tracking and other related industries.So how to design a detection model with high efficiency to approach or even replace people has become a vital problem.At present,deep learning performs better than the traditional visual algorithm in feature extraction.Therefore,based on the traditional algorithm along with deep learning face detection algorithm,this paper concentrates on invariant features,scale independence,region proposals,imbalanced samples and scale distribution with partial efficient strategies.Meanwhile,in order to accelerate the inference,we will design new convolution modules and simplify the whole network structure.Then,refining the training methods to further improve the performance.Finally,we achieve a great face detection model with better capability and faster speed by extensive ablation experiments.First of all,we implement a cascaded CNN model based on MTCNN,which further increases the width of deeper convolution channels and the size of some feature layers,and optimizes the ratio of positive and negative samples to augment the ability of detection.In addition,by merging the pooling layer to the convolution layer for direct down-sampling,the network parameters can be decreased and the inference can be speed up.Then,based on SSD,we design an anchor-based face detection model,which adds one or more strategies including dense sampling to optimize region proposals,introducing the up-sampling to modify the head layer,using the feature pyramid network to achieve scale independence,and using context module to further enhance the representation ability of the prediction layer.In addition,during training,setting different sample types to optimize the model by changing the score of IoU,and trying out OHEM or Focal Loss to solve the imbalance between positive and negative samples and improve detection performance.Finally,we realize an anchor-free face detection model,which is similar to CenterNet based on key points.By optimizing the network structure,we adopt the standard convolution in shallow layer,and the separable convolution in deep layer.Moreover,we construct a new activation function similar to the CReLU based on the Mish.Next,the convolution kernel size of 5 × 5 and the smaller prediction layer are set in some deep layers.As for the training,the 512 × 512 input images are used for preliminary training,and then the image is enlarged to 800 to fine tune the model.The experimental results show that the last one is better than the first one in detection effect,and better than the second one in detection speed.Comparatively speaking,the last one with lightest weight has more application prospects.