The thesis addresses the problem of ear-based person recognition with the aim of improving accuracy and reliability in the process of ear detection and recognition, as well as developing a pipeline that enables real-time operation using deep neural networks. The solution consists of two parts: ear detection using the YOLOv8 model and recognition employing a Siamese model, combined into a unified system that operates with a single ear image on an open dataset. We evaluated two different variations of the Siamese model for recognition on open sets. The model based on the ResNet architecture proved to be superior to the model based on the EfficientNet architecture. With this work, we have demonstrated that Siamese neural networks are suitable for ear-based recognition and that there is significant room for improvement in this area.