Patch-based Segmentation of Latent Fingerprint Images Using Convolutional Neural Network

Figures & data

Figure 1. Sample latent fingerprint images from IIIT-D database.

Figure 2. Example of convolution operation.

Figure 3. Max pooling.

Figure 4. Block diagram of proposed approach.

Figure 5. Architecture of proposed CNN model.

Figure 6. (a) 32 × 32 patches extracted from fingerprint area of a latent fingerprint image from IIIT-D latent database. (b) 32 × 32 patches extracted from non-fingerprint area of a latent fingerprint image from IIIT-D latent database.

Figure 7. (a) Segmentation result before false patch removal technique. (b) Segmentation result after applying false patch removal technique.

Figure 8. (a) Fingerprint patches. (b) Non-fingerprint patches.

Figure 9. 64 convolutional kernels of size 11 × 11 × 3 learned by first convolutional layer.

Figure 10. Result of proposed segmentation technique on latent fingerprints from IIIT-D latent database. Left column contains input images and right column their corresponding segmented image.

Table 1. Accuracy of our model on latent fingerprint patches from IIIT-D database.

	Predicted Class
Actual Class	Fingerprint	Non Fingerprint	Accuracy
Fingerprint	1022	196	83.90 %
Non Fingerprint	63	1065	94.44 %

Table 2. Performance comparison of different segmentation techniques.

Approach	Database	FDR	MDR	Avg
Ridge Orientation and frequency computation	NIST SD27	47.99 %	14.78%	31.38
Adaptive Total Variation	NIST SD27	26.13%	14.10%	20.12
K-means Clustering	NIST SD27	26.06%	4.77%	15.42
Fractal Dim & WELM	NIST SD27 IIIT-D (Good Quality)	18.7% 10.07%	9.22% 6.38%	13.96 8.23
Our Method	IIIT-D IIIT-D (Good Quality)	5.2% 4.7%	13.8% 10.5%	9.2 7.6