Most of recent advances in the field of face recognition are related to the use of a convolutional neural network (CNN) and the availability of very large scale training datasets. Unfortunately, large scale public datasets are not available to most of the research community, which therefore can hardly compare with big companies. To overcome this drawback, in this work we suggest to use an already trained CNN and we perform a study in order to evaluate the representation capability of its layers. Most of previous face recognition approaches based on deep learning use a CNN self-trained on a very large training set, taking one on the last intermediate layer as a representation and adding a classification layer trained over a set of known face identities to generalize the recognition capability of the CNN to a set of identities outside the training set. The idea is that the representation capabilities of the last one of two layers of a deep trained CNN is higher than traditional handcrafted features. In this work, starting from a CNN trained for face recognition, we study and compare the representation capability of several different layers in CNNs (not only the last ones) showing that they contain more accurate information about the face image than to believe. The proposed system extracts learned features from different layers of a CNN and uses them as a feature vector for a general purpose classifier. Moreover, we study the independence of the different sets of features used and between learned and handcrafted features, showing that they can be exploited to design an effective ensemble. The proposed approach gains noticeable performance both in the FERET datasets, with the highest performance rates published in the literature, and the Labeled Faces in the Wild (LFW) dataset where it achieves good results. The MATLAB source of our best ensemble approach will be freely available at https://www.dei.unipd.it/node/2357 “+Pattern Recognition and Ensemble Classifiers”
Alessandra Lumini, Loris Nanni, Stefano Ghidoni (2016). Deep Features Combined with Hand-Crafted Features for Face Recognition. INTERNATIONAL JOURNAL OF COMPUTER RESEARCH, 23(2), 123-136.
Deep Features Combined with Hand-Crafted Features for Face Recognition
LUMINI, ALESSANDRA;
2016
Abstract
Most of recent advances in the field of face recognition are related to the use of a convolutional neural network (CNN) and the availability of very large scale training datasets. Unfortunately, large scale public datasets are not available to most of the research community, which therefore can hardly compare with big companies. To overcome this drawback, in this work we suggest to use an already trained CNN and we perform a study in order to evaluate the representation capability of its layers. Most of previous face recognition approaches based on deep learning use a CNN self-trained on a very large training set, taking one on the last intermediate layer as a representation and adding a classification layer trained over a set of known face identities to generalize the recognition capability of the CNN to a set of identities outside the training set. The idea is that the representation capabilities of the last one of two layers of a deep trained CNN is higher than traditional handcrafted features. In this work, starting from a CNN trained for face recognition, we study and compare the representation capability of several different layers in CNNs (not only the last ones) showing that they contain more accurate information about the face image than to believe. The proposed system extracts learned features from different layers of a CNN and uses them as a feature vector for a general purpose classifier. Moreover, we study the independence of the different sets of features used and between learned and handcrafted features, showing that they can be exploited to design an effective ensemble. The proposed approach gains noticeable performance both in the FERET datasets, with the highest performance rates published in the literature, and the Labeled Faces in the Wild (LFW) dataset where it achieves good results. The MATLAB source of our best ensemble approach will be freely available at https://www.dei.unipd.it/node/2357 “+Pattern Recognition and Ensemble Classifiers”I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
