An efficient deep learning approach for 3D face detection using multiple angular 2D images

Abstract

3D face reconstruction is a useful computer vision technique for facial recognition. Accuracy decreases drastically while extracting features from 2D and moving images. To overcome this problem, we are proposing reconstruction of 3D models generated by multiple 2D angular images. Our primary approach consists of the following steps: rebuilding 3D mesh from 2D image, feature extraction, deep learning algorithm for recognition. We will be taking images of 0°, +10°, +20°, +30°, +40°,+50°, +60°, +70°, +80°, +90°, -10°, -20°,-30°, -40°, -50°,-60°, -70°,-80°, -90° angular deviations. We have compared the results from 2D architectures and 3D architectures and showed that 3D deep learning models perform better on angular images and in motion. The proposed method is time efficient and robust in nature, and it overcomes the previous limitations.