Classificação de imagens multiespectrais da mão utilizando análise por componentes principais e KNN

Abstract

Biometrics is the science that seeks to establish the identity of an individual. The process of identifying an identity involves the classification of patterns obtained from the individual's physical, biological or behavioral characteristics. Biometrics through the geometry of hand veins is a safer and more consistent method, as this pattern is unique to each person. This biometric method can be applied through images in the near infrared spectrum (NIR). The objective of this work is the classification of images in the NIR spectrum of the back, palm and wrist of hands for the identification of individuals. Principal Component Analysis (PCA), which defines the variability between image features and reduces data dimensionality, in order to extract the most important features from images. The classification was done by the k-nearest neighbors algorithm (KNN) with the Euclidean distance and the Manhattan distance using one and three nearest neighbors for each distance. The database was built through the acquisition station together with the JAI AD-080GE camera, which works with the NIR spectrum. The base contains images of 56 people, with images of the palm, back and wrist. The average of KNN hits using a nearest neighbor was 93.04% for Euclidean distance and 94.20% for Manhattan. Using three nearest neighbors, the Euclidean distance obtained 88.57%, and Manhattan, 89.82%. The average of correct answers for the back and palm of the hand obtained, on average, between 89% and 95% of correct answers, reaching 98%, while for the wrist, the averages were between 86% and 88%. The overall average of the results was 91%. The classification using the Manhattan distance was, on average, 2% better than the Euclidean distance. In many cases the two were equivalent with a high percentage of correct answers. As future work, we intend to explore pre-processing methods for extracting regions of veins from images and test such methods in other databases.