Research On Face Recognition For Anti-illumination Interference

Due to the advantages of being non-contact, initiative and user-friendly, Facerecognition system has a wide range of applications in access control systems and timeand attendance systems, especially works well in security systems in densely populatedareas and mobile payment systems in electronic commerce fields. Face recognitiontechnology need to solve the interference problems such as facial pose, expression andenvironmental illumination variation during the application. Since the image acquisitionmodule of face recognition systems usually works in a uncontrollable light sourceenvironment, illumination variation has a particularly significant effect on the facerecognition systems. Studies have shown that the difference among face images causedby illumination variation may be greater than which caused by individual facial features.And so far there has been no effective approaches can completely solve the lightinterference. Therefore research on face recognition algorithm which is outstanding foranti-illumination interference has an important significance.The approaches utilized in face recognition system to overcome the interference ofillumination variation mainly include illumination normalization, illumination invariantextraction and basis image based method. The main focus of this thesis is research onillumination normalization and illumination invariant extraction algorithms. The maincontents of this thesis are as follows:①Introduces some illumination normalization algorithms in detail, such asgamma correction, logarithmic transformation, histogram equalization andhomomorphic filtering. Then experiments about illumination compensationperformance of the methods above are given based on Yale B face database. Resultsshow that although illumination normalization algorithm can reduce interference ofillumination variation to some extent, it couldn’t achieve satisfactory performance whenstrong light exists.②Focus on illumination invariant extraction based face recognition algorithms,including MSR, LTV, LWT and NSCT based algorithms. These algorithms can reduceillumination interference existing in face images. Moreover, LTV has better capabilityof edge preserving compared with MSR. LWT even can obtain more facial informationby multi-resolution analysis of face images. NSCT is more suitable to describe textureand contour of face image compared with wavelet transform. Then classification experiments are implemented with illumination invariants of Yale B, Extended Yale andCMU PIE face databases which are extracted by algorithms above. We can also draw aconclusion that the recognition rate of NSCT based face recognition algorithm issignificantly higher than the previous algorithms.③Takes it into account that there is still facial information remained in NSCTlow-frequency sub-band, and a face recognition algorithm with fusing adaptivesmoothing and NSCT is proposed. Adaptive smoothing technology is utilized to furtherextract facial details form low-frequency sub-band, and compensate illuminationinvariant extracted by NSCT. Illumination invariant extracted by our proposed approachhas better capability of preserving facial texture and detail feature, and the proposedapproach can effectively reduce the interference of illumination variation. We appliedMSR, LTV, LWT, NSCT, INSCT and the proposed method to Yale B, Extended Yaleand CMU PIE face databases to extract illumination invariant for classificationexperiment. Results show that the proposed approach has higher recognition ratecompared with previous methods.