Fusion of CNN and LBP-HOG Features for Face Detection

International Journal of Research and Scientific Innovation (IJRSI) | Volume VII, Issue VI, June 2020 | ISSN 2321–2705

Fusion of CNN and LBP-HOG Features for Face Detection

Gopika G Das
Department of Computer Applications, Sree Narayana Guru Institute of Science and Technology, Manjali, N.Paravur, Ernakulam, Kerala, India

Abstract— Face recognition is widely used in security based applications.Even mobile phones and other such gadgets consider face as one of the most secure biometric application. It is necessary that the biometric authentication system needs to prevent sophisticated spoofing challenges. Advantages of deep learning, LBP-HOG and convolutional neural network are used in spoof detection.

Keywords— Convolutional neural network (CNN), LBP-HOG, fake face detection, face detection, face liveliness.

I. INTRODUCTION

Face anti-spoofing , as a security measure for face recognition system, are widely used due to the diversity of spoofing types. In some spoofing attacks no obvious visual cues are available to pick the genuine face images. Therefore much more generalized and discriminative features for face anti-spoofing, such as LBP and HOG are employed. These are called handcrafted features because they are designed manually. Also the features learned from convolutional neural networks are able to catch more discriminative signal in a data-driven manner. The success rate of deep convolutional neural network in the field of image classification and object recognition has attracted researchers to utilize these multi-layer end-to-end learning architectures to perform a variety of tasks. CNNs consist of many convolutional layers, followed by fully connected layers, to produce a probability distribution for the training classes.
The combination of feature fusion set of LBP and HOG are fed into Support Vector Machine (SVM) for classification process. For detection multiple feature vectors have been proposed like HOG, local binary pattern and haar like feature. One of the most popular classifier due to its efficiency and performance is SVM. The features extracted are used for training the model and this trained model is used for decision making. Convolutional neural network has achieved great success in the field of computer vision. Deep residual network is a successful improvement that breaks the limitation when trained with those containing hundreds of layers that work on extremely large dataset with thousands of classes and millions of samples, residual network is not easy to train comparing with CNN. Therefore, it is ideal to use a mixed convolutional layer followed by a residual layer can obtain the advantage of both kinds of architectures.

Read More