Research Of Face Recognition Improved Algorithm Based On LBP Features

Face recognition is one of the research hotspots in the area of computer vision and pattern recognition, with a wide application perspective. A wealth of experience and algorithms has been obtained after nearly 50 years of development. In this thesis, a research on the classical face recognition is made, with the Local Binary Pattern (LBP) studied and improved, and finally, a face recognition improved algorithm based on LBP features is proposed.Firstly, a research of the face image preprocessing algorithm is made, with the de-noising processing, geometric normalization and light preprocessing methods analyzed. Normalized standards are established for the original image rotation, scaling and tailoring. Because the face images are affected by interference pulses seriously in the process of transmission and storage, the experiment proved that the median filter has a great advantage in solving the face noise.Several common illumination preprocessing methods are studied, with experiments showing that SQI and Different-of-Gaussian (DoG) have good performance. For DoG has a great advantage in speed, this method is selected as the illumination preprocessing approach in the article.Secondly, this thesis studies the application of the local binary pattern and its variant algorithm in face recognition, with summarizes and reviews of the face recognition based on LBP made. This thesis focuses on the basic principle, evolution, and the advantages and disadvantages of LBP face recognition algorithm as well as improved face recognition algorithm based on LBP:Local Three Patten(LTP) and Three-Path/Four-Path Local Binary Pattern (TPLBP/FPLBP) algorithm. The application of several kinds of improved LBP algorithm is studied. For small posture change, this thesis puts forward the "double circle" LBP, which can further improve the LBP rotation invariance. Because LBP based on block has good recognition effect, this thesis put forward the idea of "multiple blocks+middle block", which effectively improves the problem that the information around the original block line cannot be extracted completely. Facial features are the key of face recognition, Different organs makes different contribution to human face image recognition..In this thesis, we put forward the idea of "key block" block weighted on the basis of the original block.Then, Fisherface is utilized to reduce dimension and extract FLDA features, for the LBP feature dimension is big with the use of the proposed "multiple blocks+middle block" method. The above method utilizes Euclidean distance as its classifier. The distances of the face images of one person may vary with posture and light, while the angle is the same. Therefore, this article utilizes the cosine similarity to replace the Euclidean distance. Taking the need of data dimension reduction into account, this article utilizes Cosine Similarity Metric Learning (CSML) method to replace FisherfaceFinally, experiments are conducted on Orl, Yale and Extended YaleB face databaseds. By comparing with LBP and algorithms based on LBP, it can be concluded that "double circle" LBP、"multiple blocks+middle block" LBP and "key block" block weighted LBP algorithms greatly improve the recognition rate, displaying the best performance. By comparing with PCA and Fisherface, the performance of CSML is the best.