Clemens Brockschmidt defended his Bachelor Thesis on Hand and Finger Detection based on Machine Learning
Touchless fingerprint recognition is a topic widely approached by researchers because of the ability to scan multiple fingers at the same time, ease of use, as well as hygienic factors. To initialize a fingerprint scan, the fingers to be scanned need to be segmented. Because traditional image processing approaches are overchallenged when considering the variety of skin colors, backgrounds, lighting, and other environment variables, deep learning approaches are investigatedin this work.
Two top-performing convolutional neural networks architectures are trained with data sets containing general-purpose segmentation data and specialized hand segmentation data sets. These networks are evaluated using two test datasets. The first test dataset is part of the training dataset used, while the
second data set includes composed images of hands artificially placed in front of various backgrounds.
While the deep learning approach achieves an accuracy of 77% to 97% depending on the test dataset, the traditional approach achieves 3% to 61%. The evaluation results show that the deep learning approach outperforms the traditional approach, especially in challenging environments where it is not possible to adjust the traditional approach to all environment variables.