Multi-pose Face Correction And Recognition Based On Affine Transformation

As one of the most promising biological characteristics identification technologies,the face recognition technology has developed at full speed in the last few decades. It hasthe important academic research value and commercial applications. In the actualapplication, collecting the face images is often done under abnormal conditions. It leadsto some of the face images’ postures change which often affects the accuracy of facerecognition system. To solve the problem of the pose changing, this paper focuses on thecorrection of the multi-pose face based on affine transformation and recognizes the faceimages based on improved SURF algorithm. The main research achievements include:i) Put forward a method to measure the rotation angle of the multi-pose face image.Object to the measurement problem of two-dimensional face image rotation angle, themethod combines integral projection with linear regression equation to deduce theformula of rotation angle measurement.ii) Put forward a method to correct the multi-pose face image based on affinetransformation and the principle of perspective. To solve the problem that the rate of themulti-pose face recognition is lower, the method for face posture correction is proposedaccording to the principle of perspective and affine transformation and combines with themeasured rotation angle to correct the test face image to proper frontal face image forfurther identification.iii) Recognize the face image with the improved SURF algorithm. Object to theproblem which the process of calculating the descriptor in the SURF algorithm only usea part of information around a sub-block, an improved SURF algorithm is put forwardand the K-Means clustering method is used in the process of matching feature points inorder to reduce the matching time.iv) Realized a multi-pose face correction and identification software. This papergives the framework and detailed design of the software based on the proposed correction and recognition algorithm.The results of the experiments show that the method proposed in the paper canrealize the multi-pose face image correction well and the improved SURF can not onlyimprove the rate of the face recognition effectively but also reduce the matching time.The realized multi-pose face correction and identification software lays the foundationfor further study on the research of multi-pose face recognition.