Impact of Occlusions in Face Recognition Models

Numerous researchers are looking on image distortion in face recognition systems, although some aspects of image processing have not yet been investigated. Occlusions brought on by shifting lighting and other distortions frequently make real-world face identification difficult. By applying controlled distortions, including left and right cropping, brightness changes, sketch effects, black spots, blurring, and random transformations, to a subset of photos from the CASIA-WebFace dataset, we created an occluded face dataset for this study. Using this obscured dataset, we assessed the effectiveness of cutting-edge facial recognition models such as LVFace, EdgeFace, Vision Transformer, Teacher-Finetune, EfficientNetV2M, and VGG16. According to experimental results, these most recent models exhibit remarkable resilience against a variety of occlusions and achieve high identification accuracy.In every parameter, Teacher-Finetune is the best.