Research On Finger Vein Recognition Method Based On Bag Of Words

With the rapid development of information technology, people have higher and higher requirement to the security and privacy of information. Finger vein recognition, as one of the most promising biometric techniques, has received considerable attention from researchers due to its advantages over other biometric techniques:(1) non-contact; (2) live-body identification; (3) high security; (4) small device size.However, in practice, finger vein recognition also has some problems. For example, the captured finger vein images may become blurred or have some displacement caused by illumination or finger motion. These limitations will affect the performance of finger vein recognition.In order to improve the performance of finger vein recognition system, in this thesis, we proposed some solutions. The main research contents are as follows:Since the captured finger vein images may become blurred or have some displacement, it is meaningful to extract robust finger vein feature. Since the Bag-of-Words (BoW) method has good resistance to occlusions, geometric deformations, rotation and illumination variations, which has been successfully applied in texture analysis and visual classification, the BoW method is an intuitively good candidate for finger vein feature extraction. And several researchers also have applied it in other biometric recognition (e.g., face, action, and iris). Meanwhile, we also observed that the finger vein images exhibit important regularity. Some local finger vein patterns are similar and appear repeatedly in the image. Inspired by the BoW method which can learn a small number of visual words as repeating appearance primitives, we represent these similar and repeated finger vein patterns as visual words, which are called finger vein textons (FVTs). Based on the above analysis, we use these typical FVTs to demonstrate finger vein images. Firstly, some robust and discriminative visual words of finger vein are learned from traditional base feature such as local binary pattern (LBP). These visual words are named as finger vein textons (FVTs), which can well represent the visual primitives of finger vein. Secondly, we represent the finger vein image as a finger vein textons map (FVTM) by mapping each patch of the image into the closest FVT. The FVTM can describe the finger vein effectively on a patch level.Based on the above analysis, we think the BoW model which can learn some typical FVTs is suitable for finger vein feature extraction. To better describe the feature of finger vein, the spatial pyramid representation of FVTs which reveals the global spatial layout and the local detail of the finger vein is used to make the feature more robust. The spatial pyramid representation is based on the improved BoW method (i.e., spatial pyramid matching (SPM)). Motivated by the SPM method, a new finger vein verification method which can extract robust finger vein feature with global layout and local details information is proposed.The FVTM can describe the finger vein effectively on a patch level. However, not all of the patches in a FVTM are equally useful. Generally speaking, there are many samples captured from the same finger. And we can extract a FVTM from each sample. Comparing each patch of the FVTMs from the same individual, we find that some patches are consistent. In other words, for two FVTMs extracted from two samples of the same finger, some FVTs are identical in the same patch of the two FVTMs. These patchesare called Best Patches. Instead, there are also some inconsistent patches possibly caused by the rotation or translation of finger vein, which tends to magnify an intra-class matching distance. These phenomena motivated us to use these consistent patches only for personalized matching. According to this consideration, we propose a finger vein verification method based on a personalized best patches map (PBPM). Experimental results show that PBPM can significantly improve verification performance. In addition, related results also demonstrate that the PBPM is robust to rotation and translation variations of finger vein images, which is very crucial to a practical finger vein verification system.