Semantic segmentation of delayered IC images with shape-variant convolution

Abstract

Semantic segmentation of delayered IC images pertains to the pixel-wise classification of various circuit components in microscopic IC images. It is commonly achieved by training deep convolutional neural networks (CNN) in an end-to-end manner, such as U-net and FCNs. The receptive field of the convolutional layer in the existing models is mostly invariant shape (commonly square receptive field). In the delayered IC images, the circuit components are however in different shapes/scales and could span a very wide region of the image. The context information thus may not be well-captured by the square receptive field, leading to degraded performance of segmentation. This project aims to apply shape-variant convolution, whose receptive field is related to semantic correlations, to semantic segmentation of delayered IC images for higher accuracy.