Table 1.

Comparisons of proposed and previous studies on finger-vein recognition.

Category		Methods	Strength		Weakness
Non-training-based	Image enhancement based on the blood vessel direction	Gabor filter [14,15,16,17,18,19,20,21,22], Edge-preserving and elliptical high-pass filters [25]	Improved finger-vein recognition accuracy based on clear quality images		Recognition performance is affected by the misalignment and shading of finger-vein images.
	Method considering local patterns of blood vessel	Local binary pattern (LBP) [12], personalized best bit map (PBBM) [28]	Processing speed is fast because the entire texture data of ROI is used without detecting the vein line
	Method considering the vein line characteristics	LLBP [13,29]	Recognition accuracy is high because the blood vessel features are used instead of the entire texture data of ROI
		Vein line tracking [30,31]
Training-based		SVM [38,39,40,41]	Robust to various factors and environmental changes because many images with shading and misalignments are learned.		A separate process of optimal feature extraction and dimension reduction is required for the input to SVM
		CNN	Reduced-complexity four-layer CNN [42,43]	A separate process of optimal feature extraction and dimension reduction is not necessary	Cannot be applied to finger-vein images of non-trained classes
			Proposed method	Finger-vein images of non-trained classes can be recognized	The CNN structure is more complex than existing methods [42,43]