Research On Face Detection Algorithm Based On Cascaded Convolutional Neural Networks

Face detection is an important step in face recognition.In 2001,Viola and Jones invented the VJ face detector which made it possible in practical use.However,the VJ face detector has poor performance while detecting the face in the wild.The introduction of DPM algorithm solves the problem of face detection in complex cases,but it has large cost in computation and manual marking.Thanks to the development of neural network,especially the convolutional neural network,and the realization of GPU programming,they provide a more effective technical solution for image processing,audio processing,video processing and so on.Convolutional neural network(CNN)has incomparable advantages in the field of face detection,such as self-learning characteristics,better adapting to the complex situation of the face,saving manual costs,etc.CNN is good for face detection,but it also has the disadvantage of poor interpretability.Although the face detection algorithm has made great progress,but many of them didn't work perfectly,there are still huge challenges,such as image quality,face pose and occlusion,facial expressions and illumination,real-time detection and other issues,these problems need to be solved.At present,the combination of the deep learning and the traditional method can solve the problem of face detection.In this dissertation,we design a face detection algorithm based on cascaded convolutional neural network,and the cascade structure is used to balance the accuracy and running time cost.We use full convolutional neural network(FCN)to exact candidate regions of human face in the first stage,which is more efficient than selective search,edge box and other algorithms.Combining with the NMS algorithm and bounding box regression during the whole process,we can get the more accurate of face position.In order to improve the accuracy and enhance the ability of the algorithm to distinguish the face,we try to improve the training method and optimized the training set,finally we realized the multi-task learning network.The experimental results show that the algorithm has high accuracy in FDDB data set,and the detection time is short.