Identificación biométrica de personas a partir de imágenes de la estructura vascular de los dedos usando procesamiento digital de imágenes

Abstract

This document explains the processes carried out to perform biometric identification of a person from images of the vascular structure of the fingers using digital image processing and computational intelligence. First, the image preprocessing is performed, a series of steps are followed to obtain the region of interest. The first step is edge detection using the Canny algorithm, in the second step, image dilation is performed to eliminate the gaps created when implementing the Canny algorithm. In the third step, the image is filled to complete the area of interest. After this step, those pixels added during dilation should be removed, for which XOR and AND logical operators are used. Implementing these operators leaves some spurious elements in the resulting image, and to suppress these elements, the image closure is used. Finally, masking is performed to remove areas of the image that are not useful for this problem, such as the background. The improvement of the preprocessed image quality is explained, and two techniques are presented that were used to improve the image. The first is the use of a Gaussian filter, which is implemented to remove lines that represent noise in the area of interest and can interfere when identifying veins. The second technique is the use of limited adaptive contrast histogram equalization. The histogram is used to improve image contrast so that veins are more noticeable and easier to identify. Using a convolutional neural network, the images are matched with the individuals. First, contrast is limited to highlight the vascular pattern in the images. After adjusting the contrast, a non-maximum suppression algorithm is created to find the pixels with the highest intensity in the image and represent the vascular pattern. The image is decomposed up to the fifth level of wavelet to extract more features that cannot be detected in the original image size. Then, reconstruction is used, using only the approximation images to compare which images can be used for training and validation of the convolutional neural network. After several tests, the best results are obtained using the approximation images of the third decomposition level. Finally, an accuracy of 79.86% is obtained.