Face Recognition Based On SDCNN

In recent years, face recognition is becoming a promising area in computer vision and machine learning. There is a variety of face recognition techniques, including shadow features extraction for classification, deep structure for learning, etc. As an efficient method for two dimensional face images recognition, convolutional neural network is frequently researched. However, fixed subsampling methods are applied in most convolutional neural networks and may cause lower recognition rate; furthermore, full connection between different feature map layers cannot decrease the number of parameters.In order to solve these problems, Stochastic Deep Convolutional Neural Networks (SDCNN) is proposed in this thesis to apply to face recognition task, making the pooling method during subsampling and way of connection between consecutive layers random. First, a new method for subsampling, in which the pooling value is chose randomly according to the neural activation’s energy, is proposed. The randomization process can make better use of subsampling layer in order to avoid simplification. Then, the stochastic partial and non-complete connection for feature maps convolution between different layers is achieved, thus making the connection more efficient and extracting more complementary information with fewer training parameters. Finally, a fully-connected layer and soft-max regression layer is implemented on the top of SDCNN structure with supervised training for corrected classification results.Experiments show that the proposed SDCNN can achieve better than other approaches including some classical face recognition algorithm and deep structure of convolutional neural networks under certain circumstances, which is practical for face recognition applications.