LDA-CNN: Linear Discriminant Analysis Convolution Neural Network for Periocular Recognition in the Wild

Due to the COVID-19 pandemic, necessity for a contactless biometric system able recognize masked faces drew attention periocular region as valuable trait. However, recognition remains challenging deployments in wild or unconstrained environments where images are captured under non-ideal conditions with large variations illumination, occlusion, pose, and resolution. These increase within-class variability between-class similarity, which degrades discriminative power of features extracted from Despite remarkable success convolutional neural network (CNN) training, CNN requires huge volume data, is not available recognition. In addition, focus on reducing loss between actual class predicted but learning features. To address these problems, this paper we used pre-trained model backbone introduced an effective deep model, called linear discriminant analysis (LDA-CNN), LDA layer was incorporated after last convolution model. The enforced learn so that variation small, separation large. Finally, new fully connected (FC) softmax activation added layer, it fine-tuned end-to-end manner. Our proposed extensively evaluated using following four benchmark datasets: UFPR, UBIRIS.v2, VISOB, UBIPr. experimental results indicated LDA-CNN outperformed state-of-the-art methods environments. interpret performance, visualized different layers t-distributed Stochastic Neighboring Embedding (t-SNE) visualization technique. Moreover, conducted cross-condition experiments (cross-light, cross-sensor, cross-eye, cross-pose, cross-database) proved ability generalize well conditions.