New chapter on ear recognition published in the Deep Biometrics book published by Springer. The chapter is made available as a preprint here: Read me
In the chapter we present a new ear recognition model, called COM-Ear. COM-Ear combines global and local information within a single CNN model and offers a high-level of explainability. Additionally, the model also exhibits considerable robustness to occlusions.
The chapter is joint work between Dejan Štepec, Žiga Emeršič, Peter Peer and Vitomir Štruc.
Overview of the proposed Deep Constellation Model for Ear Recognition (COM-Ear). The model is designed as a two-path architecture. The first path, denoted by 1) represents the global processing path that encodes the input image at a holistic level using a backbone CNN-based feature extractor denoted by 1b). The second path, denoted by 2) represents the local patch-based processing path which extracts features from local image patches via the backbone CNN, denoted by 2b). Local features are then combined with the PRI-Pooling operator, denoted by 2c). Global and local features are concatenated in 3) and used in the fully connected layers 4) and 5) to predict outputs and to compute losses during training.
