Research On Restoration Method Of Finger Vein Incomplete Information Based On Deep Learning

In recent years,finger-vein feature has been widely used in personal authentication due to the unique advantages of anti-counterfeiting and convenice.During capturing finger-vein images and extracting venous networks,the integrity of finger-vein images and venous networks is unsatisfactory,which affects the recognition accuracy.In this thesis,the finger-vein incomplete information restoration work includes the following two aspects.In the process of actually capturing finger-vein images,the fingers are faced with contamination;In addition,to extract the features of the venous networks,the venous networks have incomplete region during pretreatment of the finger-vein image.We deeply explore the contaminated finger-vein images and the vascular networks incomplete phenomenon aiming at establishing a finger-vein image restoration model based on deep learning framework.The main contributions are as follows:1)A network model is established to restore the contaminated area of finger-vein images.We adopt two methods to detect the contaminated area,which are based on sliding window detection and adaptive threshold detection.The advantages are compared and analyzed between the two detection methods.An encoder-decoder network model and a U-Net network model are designed to restore the contaminated areas of finger-vein images based on an intelligent restoration mechanism.2)A network model is established to inpainting the incomplete venous networks.We use two different detection methods,which are based on KMeans cluster detection and adaptive threshold detection.An encoder-decoder network model is designed to inpainting the incomplete region of venous networks.3)Experiments are implemented to verify the effectiveness of the proposed method.The experimental results show that the deep learning based finger-vein incomplete information restoration method can effectively improve the integrity of finger-vein images and venous networks,and improve the recognition accuracy.