Optimal Search Space Strategy for Infrared Facial Image Recognition Using Capsule Networks

In this paper, we propose a highly accurate method for person identification in surveillance using infrared cameras. Our model performs well when faced with the challenges of variation in view, pose, expression, scale and lighting. It outperforms the Convolutional Neural Network in scenarios where there is a continuous change in the position and translation of the targeted individual. Our error rates were 1.5 times lower than the error rates of CNNs when tested on some standard infrared and thermal datasets. We have used Local Quantized Patterns to partition people based on their genders. The people in each gender group are identified by dynamic routing between capsules. Our contribution in this paper is a new approach to filter people based on their gender and classify them using the Capsule Network. The method was tested on two infrared datasets and four visible-light-based datasets and the average error rate converged between 1%–3%.