U-Net with Dense Connections in Encoder Layers for Processing Defocused Fingerprint Images

Fingerprint features are widely used in individual identification in the field of forensic science. However, the lens of a digital camera may be in a defocused state when we take photos to collect fingerprint images, which results in blurred fingerprint images. A U-Net with dense connections in the encoder layers (EDU-Net) is proposed for processing blurred defocused fingerprint images in this paper. The EDU-Net is basically a bilateral symmetrical structure. The EDU-Net consists of the encoder module on the left, the decoder module on the right, and the “Bottleneck” module at the bottom. The input images of the EDU-Net are the blurred defocused gray fingerprint images, and the output images are the enhanced fingerprint images. There are a total of 4 encoder layers in the encoder module of the EDU-Net. Each encoder layer consists of two convolutional layers, a rectified linear unit ReLU activation function layer, and another convolutional layer in sequence. In each encoder layer of the EDU-Net, the input and the output of each convolutional layer is densely connected to the corresponding decoder layer. The encoder module is employed to extract the fingerprint feature information. The “Bottleneck” module has only one layer. This layer is used to process the deep feature information output by the encoder layer. There are 4 decoder layers in the decoder module. Each decoder layer consists of two convolutional layers, a rectified linear unit ReLU activation function layer, and another convolutional layer in sequence. Each decoder layer receives the feature maps output by the corresponding encoder layer and the upsampled feature maps output by the next adjacent decoder layer. The decoder module is used to recover fingerprint images from extracted fingerprint feature information. The loss function is composed of the pixel value loss function and the image structure loss function. The EDU-Net is experimented on the training dataset and the test dataset. There are a total of 26598 pairs of fingerprint images in the training dataset. Experimental results show the effectiveness of the EDUNet in this paper.