Attention-based Deep Learning Method for Arm Image Deblurring

Individuals could be differentiated by biometric characteristics. Arm is an emerging biometric which contains rich skin and vein distribution information. However, when acquiring arm pictures, there may exist motion blur which will cause image degradation and further deteriorate recognition accuracy. Also, different from natural scene images, vein information is hard to recover from deblurring. To solve the problems above, we present an attention-based deep learning method for arm image motion deblurring. Firstly, in order to recover more vein details, we use the near-infrared (NIR) image of the corresponding RGB arm image as a priori of vein location. Secondly, since we found more information distributes in the R channel of RGB arm images, we use channel attention blocks to further improve the recovery of blurred arm images. Our method performs better than the SOTA algorithms in terms of vein detail recovery and the accuracy of authentic verification.