Process variation band (PVB) is important for a number of lithography applications such as yield estimation, hotspot detection, and so on. It is derived through multiple lithography simulations of a mask pattern while optical settings such as dose and focus are varied. Quick estimation of PVB has been studied. A simple approach assumes optical settings for innermost and outermost PVB contour; it requires only two simulations, but the assumption of such optical settings does not always hold. We postulate that two sets of good custom kernels exist; one set for lithography simulation to extract outermost PVB contour, and the other for innermost PVB contour. Since lithography simulation can be mapped to a convolutional neural network (CNN) with kernels corresponding to convolution filters, each set can be obtained by training corresponding CNN with a number of sample reference contours. Our experiments indicate that the average intersection over union (IoU) between reference- and predicted-PVBs reaches 97% with 0 PBVs having IoU smaller than 50%. This can be compared to the state-of-art of PVB prediction using conditional generative adversarial networks (cGANs), where average IoU is only 89% with 12 PBVs having IoU smaller than 50%.