Personal Photo Album Management System Based On Deep Learning Face Clustering Technology

Nowadays,people want to save tens of thousands or hundreds of thousands of images on their computers.Along with this desire,they also want to manage their photo album based on different faces quickly.So,it is the necessity of the time to develop the tools which can fulfill this task accurately.There are many conventional techniques already used for face recognition.However,these techniques face many issues,such as postures,light,shadow,different head positions,face positions/landmarks,image size,and clarity.Conventional techniques for feature extraction have a low extraction rate,accuracy,and consume more time.There are several problems for face clustering,such as large amounts of data,a large number of classes,high variation within the class,and minimal separation between the classes.Besides,traditional clustering methods are not very effective and efficient,such as k-means and Hierarchical Agglomerative Clustering(HAC)should indicate the number of clusters.In order to solve this problem,density-based clustering algorithms are being developed.Local binary patterns(LBP)is not invariant for rotations since the positions of objects are fixed.The features size increases exponentially with the number of neighbors,which leads to an increase in computational complexity in terms of time and space.The primary purpose of this research is to propose an efficient,reliable,and scalable PAM(Photo Album Management)system based on deep learning that tackles all the issues mentioned above.The main contribution of the proposed PAM system that mainly uses the MTCNN(Multi-Task Cascaded Convolutional Neural Networks)model for facial recognition is simple,straightforward.It effectively copes with the different angles of facial expression in the natural world.Therefore,this is a reliable alternative to use powerful deep learning techniques instead of traditional facial annotation and CBIR approaches.Gradually,conventional models have been replaced by powerful deep-learning models,such as MTCNN and VGGFace2,which deal with facial recognition problems and outperforms from the existing techniques.In this way,the facial recognition task efficiently achieves the desired result for our proposed PAM system.In this research,we used MTCNN and VGGFace2 models for facial recognition and feature extraction,and the cosine distance function for similarity measurement.The cosine distance function measures the similarity between faces with a threshold of 0.4.In the experiment,the total number of 5,100 images was taken from the IMDB-WIKI data set,and these images experimented using the proposed PAM system by parts.Experimental results demonstrate satisfactory performance.It achieves a 97.28%recall rate with a precision of 92.1%and achieves 96.26%accuracy based on our proposed PAM clustering system.