Occluded Face Recognition Using Deep Convolutional Neural Network with Sparse Representation

The restricted ability to distinguish faces under occlusions is a topic that has been around for a long time and offers a unique difficulty not just for face recognition algorithms but also for humans. When compared to other obstacles, such as position variation, varied expressions, and so on, the topic involving occlusion has received significantly less attention from researchers. Despite this, occluded face identification is absolutely necessary in order to realise the full potential of face recognition for use in real-world applications. Several factors have contributed to the rapid development and improvement of facial recognition algorithms throughout the years. Researchers have researched and created a multitude of algorithms for occluded face recognition because of the unexpected aspects encountered in real-world circumstances. Yet, as a result of the epidemic, masked face recognition research has emerged as a distinct branch of occluded face identification and a pressing challenge in the field. The feature extraction stage of a convolutional neural network (CNN) is recommended to benefit from the addition of a sparse representation layer in this paper. Our objective is to improve a target network's functionality by including sparse transforms into it, all without requiring the network to perform more mathematical operations. To begin, the method that will be shown was developed by beginning with the shallow layers of a target network and then proceeding to add the sparse representation layers. Following that, the network was trained with four distinct datasets. The Difference of Gaussian Filter, also known as the DoG Filter, is a pre-processing technique that is used to minimise noise before training. The accuracy of the model that was proposed is higher than the accuracy of the other model with 98.13%.