Principal Palmprint Extraction Method Based On Deep Convolutional Network

The current principal palmprint extraction algorithm is mostly based on specific database,requiring monochromatic background or fixed postures of hands to extract ROI.If the background is complicated and the picture is unevenly illuminated,or the hand gestures are in many shapes.It usually not work well,or may needs a lot of manual adjustment for the parameters and complicated preprocessing procedure to adapt the original algorithm.For this reason,a novel region of interest(ROI)extraction method is proposed in this paper.In this paper,a small sample of image are rotated,mirrored,scale transformed and other data augmentation methods used to obtain enough training samples.Then,In order to get the mask of the palm,these images are sent and trained with the deep convolutional Encoder-Decoder network.Next,the distance transform was applying to the extracted hand mask to get the centroid of the palm and the biggest inscribed circle of the palm.The Minimum outer square of the inscribed circle is used as a mask for the palm of the hand.The palm mask is segmented into patches with size of 12*12,and the different quality versions of these patches are generated by the edge preserving guide filter,and the network is designed and trained for each version of the patches,trained the network until it can extract the corresponding features well.Then the weights of the three networks are merge at the the dense layer to fuse a new network which take these three quality patches as input and output corresponding labels.The labels determine whether each patch is dominated by the palmprint area.The experiment is done on self-built database,and its extraction accuracy is approximately at 96.2%.