POC-net: pelican optimization-based convolutional neural network for recognizing large pose variation from video

Abstract

Nowadays, face recognition using video surveillance systems becomes one of the active research topics in security domains. Security plays a significant role in everyday life for secure and sustainable developments of smart cities. The conventional techniques provide efficient recognition results only when the faces are captured with complete face images. However, they suffer to handle large pose variation images extracted from video sequences. Therefore, to deal with this issue, this paper specially designed a multidimensional face recognition model to recognize faces under multiple pose variations and angles. Three face video databases, namely facesurv database, IARPA Janus benchmark database and McGill database, are utilized for experimental evaluation. The videos of these three databases are converted into number of image frames through background subtraction process. From the image frames, the large pose variation images with different angles are identified and selected to process further. The video recorded under dynamic environment conditions diminishes recognition performance, so the image frames are processed through several preprocessing pipelines. The preprocessed images are then fed into the proposed optimal mask region-based convolutional neural network with modified short-term memory (OMRCNN-MBiLSTM) model, which learns the facial patterns present in the images more efficiently. The feature vectors learned by the proposed classifier are matched with the input face database to determine the identity of the person. With the ability to handle multiview and large pose variations, the proposed model accurately recognizes faces. The simulation result manifests the superiority of proposed model over other existing methods.