Direct Regression of Distortion Field from a Single Fingerprint Image

TL;DR

This work tackles fingerprint distortion by directly regressing a dense distortion field from a single distorted image, eliminating reliance on pose normalization and low-dimensional PCA representations. The authors design a multi-scale CNN with self-reference cues, introduce the TDF-V2Distorted fingerprint dataset, and define a ground-truth field generation pipeline to train the model. Results show state-of-the-art distortion estimation and improved rectified fingerprint matching against PCA-based methods, with favorable model size and efficiency. The approach holds practical impact for robust fingerprint recognition in the presence of plastic distortion, though future work will address noisier data and combination with prior information.

Abstract

Skin distortion is a long standing challenge in fingerprint matching, which causes false non-matches. Previous studies have shown that the recognition rate can be improved by estimating the distortion field from a distorted fingerprint and then rectifying it into a normal fingerprint. However, existing rectification methods are based on principal component representation of distortion fields, which is not accurate and are very sensitive to finger pose. In this paper, we propose a rectification method where a self-reference based network is utilized to directly estimate the dense distortion field of distorted fingerprint instead of its low dimensional representation. This method can output accurate distortion fields of distorted fingerprints with various finger poses. Considering the limited number and variety of distorted fingerprints in the existing public dataset, we collected more distorted fingerprints with diverse finger poses and distortion patterns as a new database. Experimental results demonstrate that our proposed method achieves the state-of-the-art rectification performance in terms of distortion field estimation and rectified fingerprint matching.

Figures (8)

• Figure 1: Impact of the proposed rectification algorithm on matching performance. The red and blue numbers over arrows represent the matching scores before and after rectification calculated by VeriFinger SDK 12.0 VeriFinger. Both distorted-undistorted and distorted-distorted matching scores are significantly improved after rectification (73 $\rightarrow$ 170, 96 $\rightarrow$ 236, 49 $\rightarrow$ 102).
• Figure 2: Failure cases of PCA based rectification. (a) Distortion details are lost. Green and red rectangles represent regions where orientation or magnitude information is lost. (b) Examples where finger pose is difficult to estimate accurately. The blue arrow indicates the finger center and direction extimated by yin2020joint.
• Figure 3: Statistics of the distorted fingerprint database collected by us. (a) Examples of all 10 distortion types in the database. The solid blue circle represents finger of front pose, the arrow represents the direction of force when pressing, and the dotted line represents finger of side pose. (b) Cumulative sum of the main distortion patterns in TDF-V2. The numbers on abscissa represent descending ranks of the top eight principal components in two axes. Finger pose of each fingerprint is normalized in advance.
• Figure 4: Structure of the proposed distortion estimation network. The network includes a downsampling module and a residual module for feature extraction, a spatial pyramid module for fusing multi-scale feature information, and a convolution regression module for predicting dense distortion field. For an input distorted fingerprint, the network takes the $16 \times 16$ pixel block as a unit and gives its corresponding distortion field (displacement from the input image to the rectification target).
• Figure 5: Regression error respectively of different distortion field estimation algorithms on TDF-V2_T. Color legend of bar chart and line chart is consistent. The degree of distortion is divided into seven intervals.