Face Synthesis and Partial Face Recognition from Multiple Videos

Abstract

Surveillance videos provide rich information to identify people; however, they often contain partial facial images that make recognition of the person of interest difficult. The traditional method of partial face recognition uses a database that contains only full-frontal faces, resulting in a reduction in the performance of recognition models when partial face images are presented. In this study, we augmented the database of full-frontal face images and synthesized two- and three-dimensional facial images. We designed a method for partial face recognition from the augmented database. To synthesize the two-dimensional (2D) facial images, we divided the available video images into groups based on their similarity and chose a representative image from each group. Then, we fused each representative image with a full-frontal face image using the scale- invariant feature transform (SIFT) flow, and augmented the original database with the fused images. To design a partial face recognition algorithm, we carefully evaluated the similarity between a set of video images from cameras and an image from the augmented database by counting the number of keypoints given by the SIFT. Compared to competitive baselines, the proposed method of partial face recognition has the highest face recognition rates in four out of six test cases on the widely used ChokePoint dataset, using most subjects (so-called subject group B) in the gallery. The proposed method also has recognition rates of approximately 22% to 72% on the test cases. The 2D face synthesis was found to outperform the three-dimensional (3D) face synthesis on a large subject group, possibly because the method of 2D reconstruction retains important facial features. The methods of augmentation and partial-face recognition are simple and improve the face recognition rate of traditional methods.