基于深度学习的亚表面指纹重构

Optical Coherence Tomography (OCT) is a high resolution imaging technology, which provides depth information of multilayer tissues of finger skin. An OCT-based fingerprint image is presented in the form of multiple cross sectional images (i.e. B-scan), which is quite different from the surface image of fingerprint. Reconstructing subsurface fingerprint image from multiple B-scans with high quality will not only solve the image degradation problem caused by surface imaging (e.g., altered or worn-out fingerprints, wet/dry fingers), but also provide compatible image data for minutiae based fingerprint recognition techniques. However, existing reconstruction methods are designed only for internal fingerprint (i.e. some layers of skin tissues) and simple edge detection approaches are used. Other features such as sweat gland, which exists in epidermal layers of skin tissues, cannot be extracted effectively. Meanwhile, those traditional hand-crafted feature based methods usually involve lots of parameters, which are sensitive to noises and are thus not robust. To solve those problem, this paper, for the first time, proposes to reconstruct subsurface fingerprint image using deep learning. The proposed method firstly uses HDCRes-34 as the backbone network to robustly segment each OCT-based B-scan into stratum corneum layer, ridge part and valley part, which correspond to stratum corneum layer, viable epidermis layer and papillary dermis layer. In HDCRes-34, an Atrous Spatial Pyramid Pooling (ASPP) structure is applied to increase the receptive fields of the network, so as to acquire more accurate feature representation. To further improve the segmentation accuracy, we applied a slide window to the adjacent B-scans to correct the false segmentations caused by the uneven distribution of the subsurface skin structure. The subsurface fingerprint is finally reconstructed by fusing the information of three layers by a weighted combination. In experiments, we randomly selected 4, 800 B-scans from 116, 000 B-scans collected from 73 volunteers to train the proposed segmentation network. To verify the effectiveness of the network, we compared our model with 21 classical segmentation models. Experimental results show that the proposed model can well segment the B-scans into three layers with different skin structures, with mean pixel accuracy of 0.956 and mean intersection over union of 0.873. To determine the most suitable weights for fusion, we analyzed the reconstruction results based on the number of sweat pores, the restoration potential of worn-out fingerprint surface and the definition of ridges and valleys of reconstructed images. Finally, we evaluated our method with the state-of-the-art internal fingerprint based methods using the number of features (sweat pores and minutiae) per pixel, the ridge-line density and Equal Error Rate (EER) on our own database, which were collected from 100 finger pairs of students (or officers) and 45 finger pairs (worn-out fingerprints) of labor workers. While 5.3×10^-4 ridge-line density is achieved, around 1.9×10^-5 sweat pores per pixel can be extracted from the image reconstructed using our method. 5.2% EER is achieved for our approach in our dataset. The results show that our method outperforms other reconstruction approaches in terms of the number of features per pixel, ridge-line density and EER. When worn-out fingers are concerned, the accuracy of subsurface fingerprint reconstructed by our proposed method is about 40% higher than that of 2D surface fingerprints collected by two commercial optical sensors.