Wafer Defect Microscopy Enhancement using Perceptually Motivated Super-Resolution Convolutional Neural Networks [dataset]

The silicon super-resolution (SiSR) network is a new convolutional neural network for enhancing the inspection of defects in silicon devices. This paper demonstrates how the proposed SiSR network is able to upscale microscopy images of patterns and defects by x4 in each direction, aiding inspection in small--medium scale device fabrication. The alignment cameras of a laser-writer were repurposed to capture a large dataset of patterned wafer microscopy examples at two different magnification factors, but small misalignments between the low-resolution inputs and high-resolution targets made training SiSR using this dataset challenging. Three SiSR variants were trained with different objectives. Examples and test metrics were used to evaluate the perceptual quality and reconstruction accuracy for each variant, but no variant achieved superior performance in both. SiSR-C, trained with contextual loss, generated the most perceptually pleasing defects, was the least affected by misalignments in the training set, and would be the most suitable for deployment.