Research On The Key Technology Of Face Recognition

Making a machine have the capabilities of understanding people’s emotion and knowing the outside world has been a long-term target pursued by people. The steam engine revolution happened between 1760s and the early years of 19 century makes machine replace human manual labour and liberate human’s arms and legs. The electrical revolution happened between 1870s and the early years of 20 century and the information revolution in the contemporary era en-hance people’s vision and auditory capabilities. At the same time, the complex and boring computation work can be done by machines. By now, the task is be-coming a subject to be solved by human intelligence (HI), that makes machines have the capabilities of feeling, understanding and judging like a human, help-ing people doing the complex and boring mental work so that people can have the time to do more interesting and innovating things.Face recognition is the procedure that through analyzing the probe image or probe image sequence of a person’s face, a machine can judge the identity, sex, age, and the facial expression of the person. It is currently a very active field in computer vision based on the subjects of probability theory, optimization technology, information theory, deep learning, and manifold learning, etc. In this thesis, two algorithms of face identity recognition and one algorithm of face expression recognition are proposed.1. The appearance of mobile devices such as digital camera and smart phone makes the image capturing be quite an easy task. Because the embed-ded mobile devices usually have low computation and storage capabilities, it is reasonable to transmit their captured images to the image processing servers which process the recognition task. After the servers complete the face recog-nition task they return the recognition results to the mobile device. Because the data volume of an face image is normally very huge, the bandwidth of wire-less network may be the bottleneck of data transmission. If the face image is compressed, the texture information may be lost. A novel framework of face recognition on the mobile device is proposed in section 3. In this framework, the Variable Length Dominant Feature (VLDF) vector is used as the face im-age characteristic representation. Specifically, the mobile device is in charge of the tasks such as image capturing, Gabor filtering, LBP characteristics ex-traction and variable length histogram characteristic computation. These tasks can be realized by the hardware circuits. After VLDF of the captured facial image is received by the servers, they compare it to the VLDF of every subjec-t’s face image in the gallery database and compute the distance between them. The identity of the subject in the gallery whose image is nearest to the cap-tured image is considered as the captured image’s identity. The servers return the identity to the mobile device and complete the face recognition. From the experimental data one can know that although there are 256 types of modes theoretically,80% of the pixels’modes concentrate on the 9 kinds of dominant modes averagely, only 1/6 of the modes of the Uniform LBP. The experimental results on the FERET face image database show that the performance based on the VLDF algorithms is better than those of PCA, FisherFace, LBP, and Gabor-M+FLDA. Especially on the recognition performance of face image with vari-able illuminance, the recognition rate is 0.9381, far higher than those of the four algorithms.2. A person’s face images are usually significant different on the different illuminance conditions and this difference may even bigger than the difference of different persons’face images on the same illuminance condition. Thus the face recognition is apt to be influenced by the ambient illumination. Compared to pixels’gray values, the texture orientation is less sensitive to the illuminance variations. Gabor operator can extract the textures of multiple orientations and scales from image, capturing the coarse structure characteristics. LBP opera-tor can capture the multiple orientations’ gradient information of each pixel on the image, obtaining the fine texture characteristics. Due to this complementary feature between Gabor filter and LBP operator, on chapter 4, a face image’s fea-ture representation called Local Gabor Dominant Direction Pattern (LGDDP) is proposed and this representation is used to construct a face recognition method. Specifically, firstly, the face image is filtered by Gabor filter, obtaining the re-sponding images with multiple orientations and scales. Secondly, LGDDP is constructed by coding the directions of the biggest and the second biggest pix-el value among the 8 neighboring pixels. Finally, compute the image’s spatial histogram and utilize the weighted spatial histogram similarities to process face recognition. Compared with the current published face recognition algorithms, the LGDDP algorithm has a higher recognition rate and a lower computational complexity.3. Expression is the reflection of a person’s inner emotion world. Using a face expression recognition, a machine can obtain a person’s state of mind and attitude to an affair. For example, using the face expression recognition, a machine can know whether students are interested to the contents taught by the teacher, whether the driver is in the fatigue status, whether the audiences are interested to the advertisements. So automatic face expression recognition has very high practical values. Currently the features used by the published re-search work of face expression recognition can be categorized into two classes, one is the feature based on face appearance, and the other is the feature based on the face geometrical characteristics. In section 5, a face expression recog-nition algorithm based on a single face image is proposed which combines the face image’s geometric characteristics and the appearance characteristics. Af- ter extracting the texture information around the regions of the different facial organs and constructing their classifiers respectively, the ADABOOST method is employed to combine these classifiers to form the final high performance facial expression classifier. This algorithm has the following achievements: 1. extracting the characteristics of the regions around eyes, nose tip and mouth edges respectively, combining these characteristics to form the final feature vec-tor. Compared to the method of extracting characteristics of the whole face, this feature vector can improve the classifier’s robustness to facial shape variations. 2. employing Gabor filter to extract the texture information of multiple scales and orientations, obtaining the local texture dominant direction (LTDD) distri-bution feature, using this feature to construct each region’s classifier.3. adopt-ing ADABOOST method to select the classifiers with high classifying perfor-mance, combining them with the appropriate coefficients to form the final high performance classifier. The experiments on the CK+ face expression database verify the efficiency of our LTDD+ADABOOST facial expression recognition algorithm.