近年來，為了在使用行動裝置進行交易的情形下驗證使用者身分，局部指紋辨識變得相當重要。局部指紋辨識是以小範圍指紋進行身分驗證的技術，因為在進行局部指紋對局部指紋的比對時，比起完整指紋能辨別身分的特徵數量會有所減少，發展更有效及準確的方法是必要的。因此，本篇論文結合深度學習及特徵描述子進行局部指紋辨識，以提取局部指紋中最細微的特徵。深度學習方法基於使用CNN架構的孿生網路，特徵描述子方法基於SIFT演算法，最後的辨識結果則由兩種方法所得的比對分數加權得出。 本篇論文針對各種情況進行實驗得到結果，像是不同的影像大小、不同的Epoch大小及不同的資料集。在FVC2002資料集上，DB1及DB3所得的EER約為4%，DB1及DB2所得的FRR@FAR 1/50000為6.36%及8.11%，這些結果證實本篇論文所提出的局部指紋辨識方法是準確及有效的。未來的研究可以向更高的影像解析度發展，指紋中細微的毛孔能作為特徵提升局部指紋辨識的效果，也可以使用不同的深度學習方法進一步簡化訓練過程。;Currently, partial fingerprint recognition has been considered and has become very important to identify a user′s authenticity in conducting a transaction through a mobile device. Therefore, developments to be more effective and accurate in identifying the authenticity of a user with a scanner reader that can only capture a small finger image area are needed. However, when applied in partial to partial fingerprint matching, there is a reduction in the features from full fingerprint image to partial fingerprint image. Therefore, we proposed this research using the combined method of deep learning and feature descriptors for partial fingerprint. The deep learning used in this research is based on the Siamese Network using the CNN architecture and the Feature Descriptor based on the SIFT algorithm to get minimal features from partial fingerprint. As the final result, the matching score is obtained by combining the scores from the two methods used (deep learning and feature descriptor). Then in the combination process, there is a weighting on the score obtained from both sides. The research results have been carried out on several variations of data such as image size, adequate epoch, and the type of dataset used. The results show that the proposed method by combining deep learning and feature descriptors method for the matching score evaluation in the FVC2002 yields an EER value of around 4% for DB1 and DB3. In addition, the result for FRR@FAR 1/50000 validation about 6.36% and 8.11% in the dataset DB1 and DB2. The result shows that the proposed method has good results in the implementation of partial fingerprint recognition. The development in further research can be developed using a dataset with a higher resolution. So that even though the recognition is carried out on a partial image, it still has featured in the form of detailed pores of a fingerprint and can use other deep learning methods to reduce the complexity of the training process.
