Method based on quadric surface model for fingerprint image segmentation

It is important to segment fingerprint image from background accurately, which could reduce time consumed on image preprocessing and improve the reliability of minutiae extraction. Methods for fingerprint image segmentation can be divided into two categories: at block level and at pixel level. This paper presents a method based on quadric surface model for fingerprint image segmentation, which belongs to method at pixel level. Three parameters, Coherence, Mean and Variance, are used as pixel features to determine if a pixel belongs to fingerprint part. First, spatial distribution model of pixels based on Coherence, Mean and Variance is acquired and analyzed. 200 typical fingerprint images, in which there are serious noises in background part, are selected from FVC2000 and FVC2002. The class of the pixel of these images, namely, fingerprint part or background part, is recorded manually. Coherence, Mean and Variance of each pixel are extracted and spatial distribution model of pixels is built by the use of different colors in displaying pixels of fingerprint part and pixels of background part. The model indicates that it is not linear apart and the performance of fingerprint image segmentation with a linear classifier is very limited. Second, a quadric surface formula is presented for fingerprint image segmentation and coefficients of the quadric surface formula are acquired by BP neural network using images above as test samples. Last, in order to evaluate the performance of our method in comparison to a method using linear classifier, experiments are performed on FVC2000 DB2. Manual inspection shows that the proposed method provides accurate high-resolution segmentation results. Experimental result shows that only 0.97% of the pixels are misclassified by our method, and linear classifier misclassifies 6.8% of the pixels.