Auto-Encoders Derivatives on Different Occluded Face Images: Comprehensive Review and New Results

This paper presents a novel approach for improving occluded face recognition performance using a family of autoencoders (AE) architectures. The proposed structures include four stages: image preprocessing, feature extraction using autoencoder derivatives, classification via a convolutional neural network (CNN), and evaluation on occluded, non-occluded, and unseen datasets. Three deep autoencoder variants with combinational loss terms have been introduced to extract features from images: Convolutional Autoencoder (CAE), Self-Supervised Convolutional Autoencoder (SSCAE), and Smooth Convolutional Autoencoder (SCAE). A Masked Convolutional Autoencoder (MCAE) is also introduced to evaluate the capability of our convolutional autoencoder model in reconstructing images from masked inputs. Seven public datasets have been utilized to evaluate the performance of the proposed methods: the Extended Yale Dataset B, FERET, CMU Multi-PIE, Occluded ORL, Masked LFW (MLFW), AR, and RMFRD. Occluded ORL is used to analyze the performance of the proposed autoencoder derivatives on severely occluded images. MLFWis used to test the generalization of the proposed methods as an unseen dataset for the encoder part of the autoencoder variants. The recognition accuracies of 100% for FERET, 99.89% for the Extended Yale B, 99.45% for the CMU Multi-PIE, and 94.2% for unseen MLFW, and the ability to reconstruct masked ORL with 90% of masking, demonstrate that the proposed methods achieve significant improvement in accuracy for face recognition with acceptable computational cost in comparison to the state-of-the-art. To the best of our knowledge, this study is the first work where convolutional autoencoder architectures achieve such performance on occluded datasets without incorporating any occlusion-specific design.