Research On Image Classifiction And Face Recognition Algorithm Based On Deep Learning

In recent years,researches on computer vision have achieved many remarkable breakthroughs,a series of excellent performance models is put forward.Penetrating in various engineering fields,deep learning has become the main research method of image processing,while traditional methods of machine learning has been unable to express features appropriately.In this thesis,we use convolution neural network to extract image depth features,and applies it to the fields of image classification and face recognition.In image classification,we use CaffeNet and GoogLeNet to perform scene classification experiment on Fifteen Natural Scene Categories.After study the principle and characteristics of CaffeNet and GoogLeNet respectively,we compare the two network's performance and analyze the reasons that the performance and speed of GoogLeNet are better than that of CaffeNet.Finally,we visualize the features of the middle layers of the convolution neural network and analyze the meaning of the features.In face recognition,we implement the complete process of face recognition from image preprocessing to face recognition.In the image preprocessing part,we use target detection framework Faster RCNN to detect face.At the same time,we modify the existing model of facial point detection and apply the nonlinear activation layer to the modified model,which make model better.After that,we use face alignment based on similarity transformation to reduce the impact of roll angle on feature extraction.In face recognition part,we train the face recognition network just like classification network training.We not only realize the face recognition model Light CNN,but propose an improved model called Normalized Light CNN.In the improved model,we apply normalization layer to Light CNN.The Normalized Light CNN obtain a better performance on LFW dataset.