Research And Implementation Of Illumination Evaluation And Compensation Methods In Face Recognition

This thesis addresses the problems of illumination evaluation and illumination compensation in face recognition technology. Face recognition is a major research focus in the areas of pattern recognition, artificial intelligence, this technology has wide applications. But under changing illumination conditions, recognition rate declines drastically, illumination problem is the most significant difficulty in the development of face recognition technology. The main researches of this paper include the following three parts:(1) A novel illumination evaluation method considering real-time processing requirement in video indexing applications. By utilizing appearance and grayscale information, this method can filter out input images with extreme illumination conditions that are beyond the capability of illumination compensation method afterwards.(2) An improved illumination compensation method based on Self Quotient Image. By adding a sharpening pre-processing step, more image edge information is preserved, and by adopting Gabor kernel filter, more illumination independent information is preserved.(3) Experiments and analyses of the proposed illumination processing methods. The experiments are performance analysis of the proposed illumination evaluation method, processing effect analysis of the improved Self Quotient Image method, and analysis of the face recognition system integrating the proposed illumination evaluation and compensation methods. The experiments are conducted in physical approach. In experiments, the proposed illumination evaluation method effectively filters out images with extreme illumination conditions;in addition, the improved Self Quotient Image method reveals more illumination independent information than frequency domain based method. When the above methods are applied in an Eigenface based face recognition system, experimental results show that the system achieves higher recognition rate with the proposed illumination evaluation method than without it, and the system achieves higher recognition rate with the improved Self Quotient Image method compared with frequency domain based method and the original Self Quotient Image method.