MLNN을 이용한 실시간 얼굴인식에 관한 연구

Abstract

This study proposed stable and fast face detection and identification system. First, for reliable and stable face detection, Haar-like Feature and AdaBoost were proposed. AdaBoost can calculate simple features using integral image fast. Its cascade structure can calculate them fast by removing much of the background data in initial stage with different number of features in each stage. As it needs less cost for calculation than existing approaches, its real-time detection performance is remarkably improved. The face recognition system proposed in this study obtains Eigenvector from face image set and analyses Principal Component using Eigenspace only consisting of useful vectors. It has been demonstrated through tests that it is simpler and more stable to process in comparison with approaches using geometric data or stereo image. However, the Principal Component tends to have lower recognition rate when there is change in illumination of images. In particular, when there are differences in illumination of learning images, as they may work as wrong feature points, the recognition rate may significantly lower. Therefore, this study equalized histogram to recognize who is in database even when there is change in illumination of input images, resulting in decreased change of lighting. It is not affected by changes in lighting and has maintained consistent recognition rate in comparison with PCA approach and approach using Intensity Normalization. In matching input image with model image, matching failure may occur even though real face images are successfully matched in face space where many faces are projected. To solve the problem, this study used Error Back Propagation learning algorithm, which is Multiple Perceptron learning. Using feature vectors obtained through Principal Component Analysis as input data for neutral network, this study presents suggestions for improvement of recognition performance. As a result, recognition rate of the proposed approach was 95.3%. A mean of 6.5% was improved in comparison with the existing approaches. In particular, matching failures occur frequently when there is hardly change in feature data such as rotated image of face. However, in this approach, only 2% of failure rate was found. This study obtained the optimal learning value by testing changes in recognition rates under a variety of environments. Finally, the proposed face detection system showed the best performance in real-time application, but changes in face expression and face detection in complex backgrounds should be further considered. If stereo vision is applied and face recognition is added for it to improve security and recognition rate, it can be effectively used for diverse systems.