Study On Some Issues For Auto Face Analysis And Recognition

The research of auto face analysis and recognition has important significance and application value in the field of computer vision, pattern recognition, and artificial intelligence, which can deal with the problem of face detection, face feature location, face recognition, expression recognition, and face reconstruction(2D/3D) by using computer analysis. It is widely used in science research (artificial intelligence), security (staying security in public based on face analysis), entertainment and animation (virtual face synthesis,3D face modeling, non-photorealistic face synthesis, etc), and education. This thesis deeply study some key problems of face analysis and recognition, including face feature point detection, face non-photorealistic drawing, face recognition fusing spatial and frequency blocking and face sketch recognition. This thesis obtains the following contents and innovations:1. The face recognition is not only important to research, but also practical to apply. We expound these by given three examples, such as IBM planning, netease mailbox, and the Olympic security. In the first section, we introduce the three major parts of face analysis and recognition, which are face recognition, expression recognition, and face synthesis. This thesis mainly describes the face recognition technology in detail and expounds its significance and application, and makes a more detailed discussion on the direction and mainstream of face recognition algorithm. In addition, it also summarizes the main face image databases that face recognition technology commonly used.2. The face feature point detection and localization problem is studied in this thesis. Above all, the important role of face feature point positioning in the face analysis is illustrated. It also has the vital role in some studies (face line drawings, face sketch recognition) in the rest of chapters. Aiming at the problem of the local texture model of traditional ASM is simple which is easy to make ASM into local minimum in the process of optimization, this thesis proposed some modified methods: firstly, using gray information to acquire pupil position is affected easily by illumination, least square method was used to locate pupil position, the pupil position can initialize global shape model and make most feature points close to its good position; secondly, building weighted local texture model based on YCrCb space, in view of the face skin color in YCrCb space has good clustering effect, and the areas in the positive and negative normal directions of face feature point are commonly skin or non-skin color, and we established texture models in skin color area based on skin clustering and in the non-skin color area, and established traditional texture model in the feature point area, respectively. The new local texture model fused more information in the real feature point, which also contain the information of the around area, thus it is more powerful to detect feature point. The experimental results show that detecting pupil based on least square method can be more precise. Meantime, the new local texture model can reduce the positioning error rate of ASM algorithm and improve the location precision.3. We also study the face non-photorealistic painting technology. Face line drawings is an important non-photorealistic face art form, which is widely used in entertainment, science-education and the non-photorealistic painting research, etc. This thesis proposes a new face line paints drawing method. This method combines sample learning method and local image processing method, making accurate and efficient face line drawings, and realize its vectorization. The algorithm used Canny operator to extract face contour and obtained a rough outline which including the outline of the face and hair, then used ASM to locate face features (got the precise localization of human eye, nose, mouth) and chose the right image processing algorithm or curve fitting technique to get the line processing of local characteristics, getting the precise outline of face organs. At last, we combined the rough outline and fine outline to get face line drawings. In order to drawing the line paints, this thesis proposed recursion adjacency priority vectorization algorithm. We used the VC++to develop and realize the face line drawings painting system, the system runs stably and results are satisfied. In addition, we also discussed non-photorealistic face painting based on pixel.4. According to characteristics fusion and image blocking have important role in face recognition, this thesis provided a fusing global and local features based on spatial and frequency blocking face recognition algorithm. Firstly, by analyzing two-dimensional wavelet packet transform and image group strategy, we found that wavelet packet transform based on frequency domain analysis can better satisfy the image group strategy, at the same time, the thesis also proposed the local matrix principal component analysis (LMPCA), LMPCA can build the global classifier by extracting frequency domain features and reflect face model structural features and the correlation information between different wavelet packet coefficient matrix. Secondly, by analyzing the advantages and disadvantages of traditional image blocking methods, this thesis proposed the annular blocking strategy and it can better reflect face topology information and the organ distribution characteristics. KLDA is used to extract the nonlinear local characteristics and build local classifiers. At last, global and local classifiers merge into joint classifier by weighted way. ORL and FERET face image databases are used to verify the effectiveness of the method, and the other mainstream face recognition methods are compared with it, our method has a higher recognition rate.5. Face sketch recognition is a new field in recent years, this thesis made a detailed description and survey of face sketch recognition and its methods. We researched two kinds of face sketch recognition algorithms, Firstly, we studied face sketch synthesis and recognition based on ICA. ICA makes full use of the image higher order statistics information and can be more localization than PCA algorithm. The experiment also showed that the synthesized sketch based on ICA is superior to PCA; in addition, in order to recognize sketch, people also consider the face structural information, this thesis proposed polar shape model (PSM) which can introduce face structure information to face sketch recognition by extracting the structural similarity between sketch and photo. Using the photos-sketch conversion relationship to convert photo to pseudo-sketch, then combined the sketch and pseudo-sketch extraction features based on ICA/LDA and shape projection vector based on PSM for sketch recognition. Secondly, since sketches synthesis is a more difficult problem than sketch recognition, and generally face sketch database have only a photo and a sketch for each person, thus face sketch recognition is essentially a single sample face recognition, in this study we proposed a new face sketch recognition method. This method coded face photo and sketch through the center error diffusion local binary pattern (CEDLBP), photo and sketch are projected into a middle space, for the single sample face recognition problem, we did image wavelet packet decomposition and used CEDLBP to code expand samples, and finally used PCA+LDA to identify sketch, experiments showed that the proposed method has better performance and higher recognition rate.