Research On Real-time Face Detection Method Based On Lightweight Network

As the basis of face recognition,face detection is one of the key technologies of face recognition.With the rapid development of China's economy and the increasingly mature artificial intelligence technology,face detection is increasingly widely used in security,transportation,e-commerce and other fields.The early face detection algorithm is mainly to extract effective features by manual design and train the classifier for face detection,which is sensitive to the impact of the environment.With the rapid development of deep learning,the detection algorithm based on convolutional neural network can automatically extract features with strong robustness through the training of a large number of data sets,so as to adapt to the unconstrained environment.However,the face detection algorithm based on deep learning also faces some problems: with the increase of the convolution layer depth,the neural network can obtain better performance,but it will increase the computation and reduce the detection speed.With the increasing demand of face detection based on mobile terminal,it is a challenge for face detection algorithm based on deep learning to maintain network accuracy and realize real-time face detection with limited computing resources.This dissertation comprehensively considered the detection strategy based on key point pairs and no anchor frame mechanism,and combined with the design idea of lightweight network,made some improvements to the existing work:(1)face detection method based on key point pairs and lightweight network.The existing target detection based on Anchor has a large amount of computation and requires careful manual design.In this paper,the detection strategy is improved to transform the detection of anchor frame into the detection of a pair of key points,and the corner pooling operation is introduced to determine the position of corner points.To extract the characteristics of the backbone network selection,experiment contrast study the performance of two lightweight backbone network,was proposed based on key points of and lightweight network real-time face detection algorithm,and the various and mainstream network face detection algorithm is analyzed,and verified the effectiveness of the method,realized the real-time face detection on the edge of the device.(2)face detection method based on frameless mechanism and deep separable convolution.After rethinking face detection task characteristics and studying the essential relationship between receptive field and effective receptive field and face scale,an effective receptive field was introduced to replace the sliding window and Anchor strategy,and a simple single stage face detection network with only 20 layers was proposed based on the anchor-free frame mechanism.At the same time,based on the idea of lightweight network design,a more efficient real-time face detection network based on the free frame mechanism and deep separable convolution is designed.The faster forward reasoning speed is obtained,and the real-time detection of human face is realized.By comparing with the mainstream face detection methods in many aspects,the advantages of this algorithm are shown as less computation and higher computational efficiency.The two structures based on lightweight convolutional neural network proposed in this paper are more concise,and do not need to design the anchor frame by hand,which can achieve better detection accuracy while realizing single-stage real-time face detection.