An Improved LTP Feature Extraction Algorithm For Face Recogniton

In recent years, as the key challenges of face recognition technology, the researchers haven’t find a recognition method which can overcome the changes of imaging condition, such as attitude, illumination, etc. Texture structure characteristics known as the invariant features of face image is not sensitive to the external conditions, therefore, the study on extracting the local model of face image based on the texture structure has profound significance.The method of face recognition based on local ternary pattern (LTP)can solvethe face recognition problem with the intense light or the complex conditions, itattracted much attention.But when the training sample is more, the dimension ofLTP character space will be very big, which doesnot only bring computing complexity, but also decrease identification speed of the system.. In addition, in order to choose appropriate threshold for LTP operator, we need to do a lot of experiment, which has greatly affected the training sample time and recognition rate. Through analysis and research on face recognition method based on local ternary pattern, from the texture characteristics of image and for the shortage of LTP algorithm, this article dopt the automatic threshold extraction method and dimension reduction method of (2D) PCA high dimension characteristic to improve it, specific improvement plan is as follows:Firstly, a local adaptive ternary pattern is proposed through the analysis ofLTP definition (Local Adaptive Ternary Pattern, LATP), which calculated simplyreal-time strongly and improved the training time. At the same time, according tothe process of LTP computing it proposed the arithmetic coding method of LATP,which removed some redundant information and strengthened the classificationperformance of the LTP feature space.Secondly, pointed to the problem that too much training sample in the facerecognition which led to the problem of high dimensional in feature space, this paper used two dimensional principal component analysis (2D)PCA to reduce thedimensionality of the LTP high-dimensional feature space, and extracted theeffective face features to further improved the LTP texture feature classificationperformance and robustness.Finally, It takes the comprehensive experiments on a standard face databasecombined the light pretreatment、the improved LTP feature extraction methodand (2D)PCA reduction dimensionality method, which validated these methodsimproved and enhanced LTP feature space classification performance, and theexperimental results were compared and analyzed, which indicated the validityaccuracy and feasibility of these methods.