Publication:
Deep learning based on u-net architecture for segmenting dorsal side of finger

Loading...
Thumbnail Image
Date
2024-07
Authors
Tham, Wei Wen
Journal Title
Journal ISSN
Volume Title
Publisher
Research Projects
Organizational Units
Journal Issue
Abstract
The dorsal side of fingers (finger knuckles and finger creases) contains useful features that can be extracted to identify individuals. Some individuals may have worn or damaged fingerprints or may not have fingerprints due to rare conditions or palm skin diseases. Thus, there is a need to explore an alternative method for authentication purposes. Conventional segmentation methods often encounter challenges with the intricate features in dorsal finger image such as complex textures, irregular shapes and variations in lighting conditions. To address these challenges, deep learning image segmentation is implemented. The objectives of this project are to implement U-Net Architecture for segmenting the dorsal side of finger and to evaluate and analyze its performance. VIA software is first used to label the finger knuckles and creases, producing the ground truth masks. After that, the U-Net model is developed in Google Colab. The model is then trained, validated and tested with the dataset provided. Modifications are made to the U-Net model to study their effects on the performance. The performance of the U-Net model is then evaluated and analyzed using metrics such as IoU, F1 score, accuracy and equal error rate (EER). The results obtained have proven that the U-Net Architecture is capable and suitable for segmenting the finger knuckles and creases. The U-Net model with added depth has the best performance, with an IoU of 0.69046, precision of 0.79590, recall of 0.77673, F1 score of 0.78565, accuracy of 0.77673 and EER of 0.22327. Overall, this study provides a good starting point for U-Net image segmentation of finger knuckles and creases.
Description
Keywords
Citation