Super-resolution enhancement of X-ray microscopic images of solder joints

Abstract

This study explores the application of single-image super-resolution (SISR) to enhance 3D X-ray microscopy (XRM) images for solder joint inspection. Three different voxel sizes (2 μm, 1.5 μm, and 0.5 μm) were used to scan solder joints, with the highest resolution (0.5 μm) serving as the training dataset by pairing it with bicubic down-sampled images. Two enhanced sub-pixel convolutional neural network (ESPCN) models were developed and trained to reconstruct high-resolution (HR) images. The models – ESPCN1 and ESPCN2 – were evaluated using structural similarity index (SSIM) and learned perceptual image patch similarity (LPIPS). Both models achieved high peak signal-to-noise ratio (PSNR) values of 40.01 dB (ESPCN1) and 40.33 dB (ESPCN2), demonstrating strong image reconstruction capabilities. Super-resolution models led to a significant increase in SSIM (12.0 %) and LPIPS (13.8 %) values compared to lower-resolution scans, with ESPCN1 excelling at the 2 μm voxel size and ESPCN2 showing better performance for 1.5 μm. Both models exhibited comparable results when compared to ground truth 0.5 μm scans, with ESPCN2 marginally outperforming ESPCN1 in comparison to cross-sectional evaluations. Overall, the study demonstrates that super-resolution models can enhance the quality of lower-resolution XRM images, offering comparable performance to high-resolution scans while reducing scanning time, thus proving the utility of SISR in industrial inspection applications.