Measurement ( IF 3.364 ) Pub Date : 2020-10-16 , DOI: 10.1016/j.measurement.2020.108621 Xiaoyu Chen; Qixin Wang; Jinzhou Ge; Yi Zhang; Jing Han
Non-destructive measurement of hand vein is challenging but has potentialities in many applications. Because the hand veins are under skin, the 3D annotations of hand veins are hard to obtain, and the captured images also have much noise from the skins. The traditional binocular vision methods and supervised neural networks are hard to implement in such situation. In this paper, We propose a end-to-end self-supervised binocular network (SBMNet) to compute disparities by matching pixels between the left and right images without annotations. The Region Strategy and Perceptual Loss are adopted in the training phase to improve the accuracy and the robustness to the noise. We set up the hand vein measurement system and collect simulated and real hand vein data for evaluation. SBMNet has made a successful attempt on non-destructive hand vein measurement and also has impressive results on the public KITTI dataset.