Develop of prototype system for people recognition based on ear biometrics

Abstract
Description
item.page.descriptioneng
In this work, a prototype of an ear biometric system based on Convolutional Neural Networks is designed and evaluated, thinking about the problems faced by people who cannot use the conventional fingerprint biometric system in Colombia. First, with the OpenCV Haar Cascade tool, the user's ear is extracted and a database of ninety-two users is generated, using data augmentation for later training. The characteristic of CNNs to extract features in their convolutional layers are used and transfer learning is performed with a Support Vector Machine as classifier that has the extracted CNN features as input. The CNN models used were VGG16 and FaceNet. A retraining of the VGG16 model available in Keras library was made, this model was retrained with images of ears so that it learns to extract its features. The FaceNet model developed by Google is used on its base form to get the features. These features are input to a C-SVM classifier, the SVM hyperparameters are adjusted with Sklearn Grid-Search technique, the CNN models use different SVM hyperparameters. Python scripts are developed to implement the proposed models, such as user enrollment, classifier training and the use of the proposed system. After having the algorithms ready, tests were made to evaluate their performance with different techniques such as Sklearn cross-validation to figure out the accuracy of the models, the False accept rate and False reject rate metrics, and finally the ROC curve for biometric systems to get the performance of this prototype system.
Keywords
Citation