Smile Recognition Based On Gabor Feature And Deep Auto-encoders

In the field of Artificial Intelligence and Machine Learning, feature extraction is one of the key technologies of intelligent human-computer interaction. As a branch of Face Recognition research, feature extraction and classification of facial expressions are becoming more and more popular and it has been an extremely active subject. To some extent, the research of facial expression is based on the research of face recognition technology, while on the other hand,the research of facial expression can also accelerate the process of face recognition research.Recently,although facial expression recognition have made certain progress,there are still many problems to be solved. Especially for those expression images which span a wide range of imaging conditions, such as variability in illumination, shelter, posture and so on. These factors may make expression recognition difficult to study.Smile expression is the most important expression of human expression,and it is also the most common expression. With the development of pattern recognition, improvement in living standards and the popularity of consumer electronic products, smile recognition is receiving increasing attention. Traditional smile recognition methods mostly focus on constrained smile,researcher do their experiments on some special constrained public library. Their experiments and methods always show higher recognition rate,but when tested in real scenes,the recognition rate drop sharply. So how to improve the robustness of smile recognition methods, make smile recognition technology overcome these influence factors: illumination, shelter, posture etc. and achieve the stage of practical application are the problems to be solved urgently.To improve the effectiveness and robustness of smile recognition method in real scenes,this paper mainly focus on those images in unconstrained database to recognition. Based on deep learning theories, this paper did some studies on smile recognition, it constructed deep neural network model as well as studied the effectiveness of fusion feature of Gabor and deep neural network model. The main contents of this paper are as follow.1. Many kind of auto-encoders has been studied. In a broad sense,auto-encoder is a network with three layers. By means of constantly back-propagation fine-tuning to minimize the errors between input and output, auto-encoder can reconstruct originated input in the output. In a narrow sense, auto-encoder is also a restricted Boltzmann Machine. The auto-encoders we studied arefrom the broad sense, and all kinds of auto-encoder variants are regularization forms of common auto-encoder.2. Smile recognition based on deep auto-encoders has been studied. Most known deep network models are stacked by some simple and same base model. In order to study the feasibility and effectiveness of the different base model formed deep neural network for smile recognition,this paper used a deep neural network which is stacked by contractive auto-encoder and de-noising contractive auto-encoder for smile recognition and classification. The experimental results show that our deep model is feasible,and the model we constructed is better than the simply stacked contractive auto-encoders and stacked de-noising contractive auto-encoders.3. Based on Gabor feature and deep auto-encoder. This paper discussed two fusion features of Gabor and used the fusion of Gabor features as deep auto-encoder model’s input. Some details of algorithms implementation are described in this paper. We studied the effectiveness of the deep neural network which take the fusion of Gabor features as input, and compared the deep auto-encoders when they under different feature input: one is the original pixel images, and the other is Gabor features or the fusion of Gabor features. Experimental results show that the fusion Gabor features as the input of deep network work faster than the pixels images input, as the same time, the fusion Gabor feature, to some extent, is more advantageous to the limited real environment face image recognition.