چكيده به لاتين
Deep learning can be considered as a new chapter in the field of machine learning. Deep learning, or deep neural networks, have caused significant advances in various fields, including image processing and machine vision.
The results obtained from deep face recognition systems have improved the performance of conventional methods. Nevertheless, lack of methods with specific conditions such as Multi-View Face Recognition is the reason of researches in this area. On the other hand, one of the important challenges in the field of machine vision is to recognize of three-dimensional objects such as faces. Hence, faces as a three-dimensional objects should have three-dimensional algorithms. Three-dimensional cameras are expensive, it takes a long time to create a three-dimensional face model and shortage of three-dimensional face datasets are the reasons that researches encourage to employ two-dimensional face datasets.
In this thesis, we present an architecture with convolutional neural networks which achieves information from multi-view face images and aggregates them into a single layer called, aggregating layer. In this thesis, we propose a method for recognizing faces by a set of two-dimensional images which takes advantage of the benefits of the two-dimensional and three-dimensional face recognition methods and reduce any disadvantages.
The purpose of this method is to increase the accuracy of recognizing faces of individuals by set of images based on view angles. Another goal is to find that the integration of which views conclude the best results. Experiments indicate that the proposed method outperforms common face recognition method.
Keywords: Deep learning, deep convolutional neural network, multi-view face recognition, multi-view representation