Fully convolutional networks for void segmentation in X-ray images of solder joints

Whether in sensory or illumination applications, optoelectronic components are an essential part of our everyday life. To actually include them in smartphones, cars or other products, they need to be soldered onto the surface of a printed circuit board via surface-mounted technology. Hereby, the solder joint is formed by remelting the solder paste that was printed onto the board before the optical component was mounted. During this process, the evaporating flux causes porosities (voids) filled with gas that is caught in the solidifying alloy. Voids influence the thermal and electric properties of a solder joint and hence reduce its reliability. Because the solder joint is embedded between the device and the board, non-destructive X-ray inspection is used to visualize voids. However, the superposition of various structures in a noisy image acquisition process renders the semantic segmentation of solder joints and voids in X-ray images difficult. To the best of our knowledge, there is no method for automatic void segmentation in flat solder joints based on X-ray images aside from this work. We develop a fully convolutional network for pixel-wise classification of X-ray images and show, how our contributions enable automatic void inspection of soldered structures without protracted X-ray tomography of flat samples, so-called laminography.