A Software-Hardware Codesign of A Deep Learning Based Fingerprint Recognition

• Main Authors: Hsiang-An Hsieh, 謝翔安
• Other Authors: Shiann-Rong Kuang
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/2spaw5

碩士 === 國立中山大學 === 資訊工程學系研究所 === 107 === Due to the recent progress in hardware development, the growth and application of artificial intelligence are increasing rapidly in the world. The broadness and powerfulness of the artificial intelligence would be able to complete multiple tasks and solve various problems for the humankind. In the past, fingerprint recognition systems were operated by computers with complex algorithms. With the improvement in the artificial intelligence techniques, more and more people are performing fingerprint recognition through deep learning. Since recognizing fingerprints with a single CNN is a great challenge, the process of fingerprint recognition is usually divided into the following four steps: capturing, preprocessing, feature capturing, and feature comparison. In the recent literature, the techniques of these steps were often proposed individually, but the entire fingerprint recognition system was discussed rarely. In this thesis, a thorough architecture of the fingerprint recognition system is being proposed. Integrating the representative techniques from the above-mentioned steps and designing the hardware circuit for convolutional function, this proposed architecture would speed up the process of fingerprint recognition. FingerNet is adopted in the feature capturing step to discover the features and output the information. After the information is transformed into the format certified by the ISO, SourceAFIS is used to compare the correlation between two fingerprints. If the comparison score exceeds the Threshold, the two fingerprints are considered to match. In order to leverage the practicality, the feature comparison step is altered to a one-to-many system. Different datasets are used to verify the proposed system, and the accuracy rate can achieve 93% with the dataset of our lab and 98% with the dataset from the Fingerprint Verification Competition.