Abstract
In real-world scenarios, the quality of identification is affected by lighting conditions, changes in facial expressions, spatial positioning, and various other external factors, which further complicate the image processing pipeline. Currently, research in this field is advancing along two main directions. The first is improving face recognition algorithms to increase robustness against partial and full occlusions. The second involves using image reconstruction technologies to restore missing facial regions, followed by recognition. The paper analyzes existing datasets and processing methods for occluded faces, comparing face recognition approaches and neural network-based occlusion handling techniques. Modern deep learning-based solutions and traditional methods are compared, highlighting their advantages and disadvantages. It is demonstrated that a model based on the Self-Attention mechanism can enhance robustness to partial facial occlusion by actively considering the overall image context. The conclusion discusses the main challenges, unresolved tasks, and promising directions for the development of face recognition technologies under occlusion conditions.