Void and solder joint detection for chip resistors based on X-ray images and deep neural networks.

Chip resistors are the most common and most frequent components in all electronics. Voids are formed by outgassing during solder reflow. Excessive voids influence the electrical and thermal conductivity of the solder joint and hence reduce its reliability. Therefore, in this study, we propose a scheme combining Residual Multiscale Skip Connected Net (RMSC-Net) and Recurrent Convolutional Network (RU-Net) to accurately detect the voids and solder joints of the X-ray images of the chip resistor. We first develop an RMSC-Net for void segmentation of X-ray images. RMSC-Net combines residual with multiscale skip connections. Residual learning solves the problem of gradient dispersion by learning the difference between the target and input values. The proposed multiscale skip connection combines multiple high- and low-resolution feature maps to realize multilayer feature perception of voids. Then, we deploy RU-Net with recurrent convolution modules to apply the chip resistors' solder joint segmentation. Experimental results demonstrate that the overall performance of RMSC-Net and RU-Net is better than other methods in detecting voids and solder joints of three different types of chip resistors. RMSC-Net achieves average SE gains of 2.70 and 2.25 points over UNet++ and FedDG, respectively. The average PR of RU-Net is 0.34 and 0.61 points higher than that of UNet++ and FedDG, respectively. • A detection scheme that combines RMSC-Net and R-UNet to detect voids and solder joints in chip resistors is proposed. • A RMSC-Net is developed for micro-voids segmentation of CT images. • We present multi-scale skip connections to fuse different levels of features. • The residual is introduced to solve the problem of gradient dispersion or gradient explosion. • R-UNet with recurrent convolution modules is adopted to detect the edges of solder joints accurately. [ABSTRACT FROM AUTHOR]