Photo Album Management System Based On Face Recognition

Nowadays,more and more photos are taken with smart phones,digital cameras,etc.,and the number of photos in the album directory on the computer is also growing rapidly.The management of these digital photos becomes indispensable.Lots of photos in common family albums are themed with people.These photos are usually displayed in chronological order.But when user want to view all photos of a person,it’s inconvenient to browse and hard to find.And it would take a lot of time and effort to organize these huge numbers of photos manually.In order to solve the problem that family albums are difficult to manage,this thesis designs and implements an album management system,which is based on technologies such as face recognition,B/S(Browser/Server)architecture,database,Python and Java Script.Face recognition related technology is the key part of this system,which mainly includes face detection,face alignment,face feature extraction,face clustering and classification.For the face detection step,the performance was compared among Haar cascade classifier,face detector based on HOG feature,face detector based on CNN model,and detector based on Center Face model.The detector based on HOG feature was chosen as the face detection method in this system.For the face alignment step,the performance was compared between the 5-point landmark model and the 68-point model.The faster 5-point model was used in this system,whose average time was 1.9 milliseconds per face.For the face feature extraction step,the Res Net-based face recognition model implemented by dlib was tested,which encoded the face into a 128-dimensional feature vector.Its average time was 0.65 seconds per face(without GPU acceleration).For the face clustering step,the performance of DBSCAN(Density-Based Spatial Clustering of Applications with Noise)and CW(Chinese Whispers)clustering algorithms on different face datasets was investigated.The results show that the clustering performance of these two algorithms has a great relationship with the choice of parameters.Improper parameter values will lead to very poor performance.In the scenario of the family album management system,this thesis proposes a new method that can adaptively determine the parameter values of the DBSCAN and CW algorithms.This method is based on a fact that in most family albums,a person’s face does not appear twice in the same photo,that is,multiple faces appearing in a photo belong to different people,which should be clustered into different clusters.This fact can be used as a constraint to evaluate the effectiveness of the clustering algorithm.A good clustering result should not violate this constraint(group photo constraint)too much,and the Violation Rate is defined accordingly.By applying this method to the parameter estimation of DBSCAN and CW algorithms,the calculated parameter values have achieved very good results on the datasets used for testing.After some comparison,DBSCAN was used in this system as it is faster than CW and can remove noise.For the face classification and recognition step,the influence of parameter values on the KNN(k-Nearest Neighbor)classification algorithm was tested.The method to identify an unknown face using the nearest known face was proved and used in this system.The functions of the album management system implemented in this thesis include browsing photos with pagination,automatically grouping photos by people,displaying people(group)list,interactively merging or removing people,and interactively modifying the relationship between photos and people.On the self-built family album dataset and a subset of the face image dataset Face Scrub,the above functions were verified and passed.The average precision of automatic grouping photos by people is 99.2%,and the average recall is 90.1%.