Robust feature extraction in fingerprint images using ridge model tracking

This paper addresses the problem of feature extraction from low-quality regions in fingerprint images. Features such as minutiae are difficult to detect automatically when high levels of noise are present, as in wet, dry, or latent prints. The approach presented here is a novel application of Bayesian filtering to the problem of ridge tracing. A friction ridge is followed by recursively estimating posterior distributions representing the direction of each new step along the ridge. This approach benefits from previous state information when in an area that exhibits low ridge clarity, causing ridge-flow estimation to be unreliable. The new technique has been tested experimentally using a database of 880 grayscale fingerprint images with varying quality. The ability of the proposed method to detect features more reliably is confirmed by a reduction in Equal Error Rate of 2.1% and 2.5% over two traditional methods. In addition, the False Reject Rate was reduced by 11.1%, at a False Accept Rate of 1%, for a group of low-quality images. These results demonstrate a significant improvement, as compared with previous techniques, in the ability to process low-quality fingerprint images.