Person identification in low resolution CCTV footage using deep learning

Abstract

The advancement in camera technology, high speed connectivity, and availability of huge information over media resulted into developing better algorithms for person identification. However, the recognition of a person in low resolution cameras like CCTV is still a challenging problem. This paper aims at identifying a person in low resolution image captured by webcam or in various frames of CCTV footage by using deep learning convolutional neural network (CNN). The proposed system uses facial image of a person for this task. The designed CNN pipeline has six convolution layers, one flattened layer and two fully connected layers. Total 6667 images of 62 subjects are used for training and validation of CNN framework with 500 epochs. The designed CNN attained 99.99% and 98.45% of training and validation accuracy, respectively. The network was tested on 1599 test images and was able to achieve a testing accuracy of 96.03%. The proposed CNN framework is also tested on low resolution face recognition benchmark dataset (TinyFace) for which it achieved 94.55% testing accuracy.