Recurrent Network-Based Face Recognition Using Image Sequences

In this work, we propose a novel method for face recognition with large pose variations in image sequences using a Cellular Simultaneous Recurrent Network (CSRN). The pose problem is still a daunting challenge in face recognition. If the image sequences are obtained from different viewpoints in a surveillance type of application, the face recognition rate drops significantly. We formulate the recognition problem for face image sequences with large pose variation as an implicit temporal prediction task for CSRN. Further, to reduce the computational cost, we obtain eigenfaces for a set of image sequences for each person and use these reduced pattern vectors as the input to the CSRN. The CSRN is trained by this pattern vector, and each CSRN teams how to associate each face class/person in the training phase. When a new face is encountered, the corresponding image sequence is projected to each eigenface space to obtain the test pattern vectors. The Euclidian distances between successive frames of test and output pattern vectors indicate either a match or mismatch between the two corresponding face classes, We extensively evaluate our CSRN-based face recognition technique with 5 persons using publicly available VidTIMIT Audio-Video face dataset [1]. In order to verify the performance of the CSRN, we also implement an Elman neural network for comparison. Our simulation shows that for this VidTIMIT Audio-Video face dataset with large pose variation, we can obtain an overall 65% (for rank 1) or 75% (for rank 2) face recognition accuracy better than the 55%(rank 1) recognition accuracy of Elman neural network.