Machine learning approach for face recognition from 3D models generated by multiple 2D angular images

Abstract

We propose machine learning approach for face recognition from 3D models generated by multiple 2D angular images that recognizes faces from multiple angle of a 3D face model. Though, many works on identifying faces from 3D have already been done, there are many spaces to update, improve and contribute more features on previously done researches. However, this research includes SFM algorithm which is a combination of SIFT detector, Approximate Nearest Neighbors (ANN) algorithm and RANSAC algorithm to reconstruct 3D from multiple RGB images. Again, it includes AdaBoost Learning algorithm which was used to train model to recognize faces. Besides, we used Local Binary Pattern Histogram (LBPH) which is an effective texture administrator, marks the pixels of a picture by thresholding the area of every pixel. Finally, the System successfully recognizes faces which are deviated up to 60°angular deviation respectively to left and right (total: 120°). Additionally, it gives an accuracy of 80% to 100% depending on angular deviation of up to from 0°to 60°. Nevertheless, the rate of accuracy of our proposed system is reversely proportional to the Angular Deviation.