Abstract
Fingerprint, one of the most popular and acceptable biometric features, is widely used in all kinds of fields. Fingerprint recognition technology has been researched for decades. However, even the state-of-the-art recognition approach, its performance depends greatly on the fingerprint quality. If we can detect the defective or low-quality regions in fingerprint images, we can repair them or downgrade these regions to improve the recognition performance. In this thesis, we propose a system for defective region detection from fingerprint images based on fully convolutional network (FCN). This system can detect defective regions which are too dry, too wet, or belongs to the background, in the pixel-wise level. The proposed region detection algorithm takes advantage of the high classification accuracy and great learning power of the deep neural network. Combined with the feature of FCN for pixel-wise prediction, we can accomplish the task of defective region detection. With pre-training on fingerprint patches and the defective fingerprint synthesis approach, we develop a FCN model with satisfactory performance on the fingerprint segmentation task. At last, our experiments show that the proposed FCN fingerprint segmentation model can be used to improve the fingerprint recognition accuracy by integrating it with the existing fingerprint recognition system.