FPD-M-net: Fingerprint Image Denoising and Inpainting Using M-net Based Convolutional Neural Networks

Abstract

Fingerprint is a common biometric used for authentication and verification of an individual. These images are degraded when fingers are wet, dirty, dry or wounded and due to the failure of the sensors, etc. The extraction of the fingerprint from a degraded image requires denoising and inpainting. We propose to address these problems with an endto-end trainable Convolutional Neural Network based architecture called FPD-M-net, by posing the fingerprint denoising and inpainting problem as a segmentation (foreground) task. Our architecture is based on the M-net with a change: structure similarity loss function, used for better extraction of the fingerprint from the noisy background. Our method outperforms the baseline method and achieves an overall 3rd rank in the Chalearn LAP Inpainting Competition Track 3−Fingerprint Denoising and Inpainting, ECCV 2018.

Using a Patch-Wise M-Net Convolutional Neural Network for Tissue Segmentation in Brain MRI Images

“…To alleviate these problems effectively, the M-net architecture was proposed for segmenting brain MRI scans. The M-net [20] architecture has two side paths (left and right legs) and two main encoding and decoding paths, which helps it learn better features. Furthermore, the two side paths aid in learning finegrained details from brain MRI scans.…”

“…In this study, we performed multi-class segmentation to label every pixel in an MRI scan as one of the four classes: background, GM, WM, and CSF. To this end, the conventional M-net architecture [20] was modified such that the final layer produces a binary map for each of the four classes, rather than a single binary map representing foreground and background pixels. Input ground-truth segmentation maps are also converted into multi-channel binary segmentation maps for each class.…”

“…However, because it uses trainable up-sampling, U-net must learn many more parameters and is much slower to train than SegNet, which does not capture multi-scale information as effectively as U-net. Adiga et al [20] proposed an end-to-end multi-class deep network called M-net with deep supervision functionality. However, the Mnet architecture has issues related to limited memory for high-resolution input images because the number of feature channels increases over the resolution of the input images at each stage of down-and up-sampling, thus resulting in a more number of parameter values to store.…”

“…Our proposed method is compared with SegNet [18], U-net [19], and M-net [20], which are tested with the parameters shown in Table 3. A total of 50 subjects are used for experiments, where 30 subjects for training and 20 subjects for testing are used, respectively.…”

“…The SegNet [18] passes pooling indices to the up-sampling layers from the encoder to the decoder. For Unet [19] and M-net [20], the decoder section mirrors the encoder with up-sampling layers concatenated with feature maps from the encoder at the same feature map level. The final output is computed using 1×1 convolution with the Softmax activation function.…”