| In recent years,the research on deep learning has been increasing.Convolutional neural network is one of the many deep learning algorithms.It has the characteristics of sparse connection and weight sharing.This paper summarizes the development history and the latest research results of deep learning at home and abroad.It briefly introduces the concepts and algorithms involved in artificial neural networks and classical convolutional neural networks.It is applied to the face detection and face landmark detection.The innovation of this paper is as follows:1.A novel face detection algorithm based on joint convolution neural network classification for candidate region generation is designed.By comparing three candidate region generation algorithms and improving the structure of convolution network,an optimal face detection method is verified.2.On the basis of the original concatenated convolution network,three different face key point detection algorithms are designed.The experimental results are compared to obtain a method with better robustness and lower detection error.3.Based on face detection and facial landmark detection,we put forward an evaluation method for students' class concentration.The main contents of this paper are as follows:We propose a series of strategies to improve the original network structure:improving Dropout method,using PReLU activation function,adding LRN and BN layers to the network structure,splitting the large convolution kernel.the accuracy of improved network classification increased by 25%.Aimed on the localization of facial five key points,the localization task is divided into two parts:feature extraction and regression localization,and we design a cascaded regression deep convolution network method.Compared with traditional DCNN algorithm,it has a great advantages.It improves the accuracy rate of facial landmark detection,reduce the average failure rate.We carry out experiments to verify the robustness and practicability of the method. |
PDF Full Text Request