Carreira-Perpiñán, M. Á. (1995): Compression neural networks for feature extraction: Application to human recognition from ear images (in Spanish). MSc thesis, Faculty of Informatics, Technical University of Madrid, Spain.


Outer ear (pinna) images are proposed for human recognition, based on their discriminant capacity (only outperformed by fingerprints) and their small area and variability (compared to face images). Compression neural networks are used to reduce the dimensionality of the images and produce a feature vector of manageable size. These networks are 2-layer linear perceptrons, trained in a self-supervised way (we used backpropagation and quickprop). A theoretical justification of the training process is given in terms of principal components analysis. The approach is compared with standard numerical techniques for singular value decomposition. A simple rejection rule, based on the reconstruction error, for recognition is proposed. Additional comments about the robustness of the network to image transformations and its relation to autoassociative memories are given.


Face processing, personal identification, pattern recognition, artificial neural networks, image compression, image segmentation, autoassociative memories, Karhunen-Loéve transform, principal component analysis, ear database.


This is probably the first work that introduces the idea of ear biometrics for automatic personal identification. Unfortunately, I regret that only the Spanish version of this thesis is available, due to an annoying requirement of the Technical University of Madrid that all theses must be written in Spanish. So much for information accessibility...


[Image (GIF 16K): ear schematic] [Image (GIF 4K): ear database]

Miguel A. Carreira-Perpinan
Last modified: Sat Jun 25 23:52:22 PDT 2016

UC Merced | EECS | MACP's Home Page