Large Scale Face Clustering Based On Convolutional Neural Network

Face recognition is an important biometrics recognition. It has been widely applied in various fields such as finance and security. In order to improve the rapidity of the extending face retrieval system, we propose a large scale face clustering algorithm based on Convolutional Neural Network (CNN). In this way, face database with large scale is divided into several smaller databases whose samples are more similar to each other and the retrieval range is narrowed.How to extract efficient facial features and choose a fast clustering algorithm with high accuracy become key technologies of large scale face clustering. We aim to distinguishing the individuals dramatically and reduce the impact caused by illumination, posture, expression, obstruction, etc. Different from the general clustering algorithms, a suitable algorithm for large-scale face clustering needs to be highly accurate and of low time complexity. Therefore, we construct a CNN to obtain effective facial features, and furthermore compare the performance of classical K-means and modern CFSFDP (Clustering by Fast Search and Find of Density Peaks).The main contents are as follows:1) A CNN for feature extraction is developed. Compared to the traditional algorithms, CNN relies less on experience and fully considers the complexity, nonlinearity as well as high order of the face. The results show that features extracted by CNN are robust against illumination, posture, expression, obstruction, etc.2) We use K-means++ to improve the initialization of clustering centers and design an appropriate index to estimate the value of k. Finally, we reduce the possibility of being trapped in local minimum and overcome the drawback of prescribing the value of k.3) MSRA-CFW database of large size and complicated variability is applied to test the clustering performance. In addition, Rand Index, entropy, and F1 measure are further used to evaluate the clustering quality. To be more intuitive, the visualization of the confusion matrix is implemented to reveal the clustering results.4) We adpot CNN, PCA, HOG features and K-means, advanced K-means, CFSFDP clustering algorithms to design the experiments. The results suggest that K-means++ based on CNN feature to be the best. Moreover, it only needs to estimate the value of k for once and has less computing time.