Researches On Face Recognition Via Compressed Sensing

As one of the hottest research directions in the field of computer vision andbioinformatics, face recognition has provided powerful technical support for theinformatization process the core applications in public security and human-machineinteraction, which has been capturing special attention from both academia andindustry. Having devoted much effort and spent a large amount of funding on facerecognition, the researchers are still not satisfied with its slow development in the pastdecades. Even today, further development of face recognition is still seriouslychallenged by complexity of illumination, variability of pose and expressions, andrandomness of occlusion.Most of the existing face recognition algorithms are based on the classicalstatistical learning theory, which has been proved to be effective in solving lowdimensional problems where sufficient training samples are available. However, theclassical statistical learning theory cannot well handle problems with highdimensional data due to the characteristic of face images. Meanwhile, the number ofcollected training samples is severely restricted in real applications. Due to the aboveaspects, the classical statistical learning theory is not very suitable for face recognitionapplications. In 2006, Donoho and Candes proposed a novel framework namedcompressed sensing (CS). This framework has aroused another upsurge in facerecognition since it's introduced into the face recognition area, and one of the mostoutstanding algorithms based on CS is Sparse Representation-based Classification(SRC).Compared with most existing algorithms, SRC implements statistical inferenceby directly exploiting the sparse distribution of high dimensional data, which canhandle the curse of dimensionality effectively. Moreover, SRC implements facerecognition via image pixel values and avoids information loss thanks to thepre-processing procedures.SRC requires exact alignment between each test image and training images.Nevertheless, variation of poses and expressions leads to the error on alignment andthus the SRC's performance may decrease. This fact severely restricts SRC'sgeneralization ability to real-world face recognition problems. This dissertationfocuses on the research of CS-based face recognition algorithms that handle pose variations and are robust to expression variations. The main contents and innovationsof this dissertation are as followes:1. Proposed a pose-robust face recognition algorithm via part-based Sparse Representation (PSRC). In PSRC, each image is represented by a set of local regions first, and then a affine transformation model is used to model the pose variation, the parameters of the affine transformation are estimated between each test image patch and the best-matched training image patch to regulate pose variation. Experimental results demonstrate that PSRC can handle large pose variations compared with existing algorithms.2. Proposed a global optimization algorithm to automatically find the optimalaffine transformation parameters. To improve estimation accuracy of these parameters, the initial values of parameters in each patch are first roughly estimated via minimizing reconstructed error. More accurate initial values mean smaller probability of being trapped into the local optimum and more robustness of the algorithm. Experimental results demonstrate that estimation of initial values improve performance of PSRC considerably.3. Proposed a facial expression-robust face recognition algorithm based on compressed sensing. The shape model is used to achieve the alignment between face image with expression and registration images. Then, dictionary is formed by front-view registration images to achieve the sparse representation of each test image,and then the identity of each test image is recognized. Experiments demonstrate that under facial expressions such as smile, frowning and surprise, the proposed algorithm achieves good recognition results.