August_EDFA_Digital

edfas.org ELECTRONIC DEV ICE FA I LURE ANALYSIS | VOLUME 24 NO . 3 20 RESULTS AND DISCUSSIONS Layout to SEM Image Translation. To quantify the quality of the synthesis process, the evaluation protocol used in traditional image-to-image translation works [14] was adopted, but in a slightly different way. First the synthetic images were binarized by performing the same binarization technique that has been used to generate layout images from cleaner SEM images. A comparison was made to see how well the predicted binarized image matches the input layout. The intuitionbehind the task is if it’s possible to synthesize realistic images that correspond to the layout, then the samebinarization technique should give images that are not so out-of-bound compared to the input layout. Table 1 shows the calculated segmentation accuracy. For both mean intersection-over-union (IOU) and mean dice coefficient score, this method performs significantly well. To the best of the authors’ knowledge, this is the first work on the SEM image synthesis from layout images of 28 nm node technology. Synthetic Cell Image Generation. Following Ref 12, the Jensen-Shannon-Divergence (JSD) method was adopted for analytical evaluation of the synthesized image. The JSDscore for the synthesized image is 0.06. The JSD score for the generated data comes out as 0.06. The lower JSD value signifies that synthetic data distribution is not out-of-distribution from the original data. Different modes of synthetic images are presented in Fig. 7. Inspecting the visual data found that MSGAN has produced a diverse set of synthetic images. But a lower JDS score also signifies that the generated data is not diversified enough. The main reason behind this is the data scarcity and not having enough data for each of the classes. Incorporatingmore images of different classes in the dataset may generate diversified images in the future. Cell Classification. The classification model has been tested with the limited number of images available, achieving very high accuracy (99%) on the test set. As the data is not diverse enough, the network learns dis- criminative features easily. Future work will incorporate a more diverse set of data for a robust generalization of the network. SUPERVISED FEATURE EXTRACTION AND SYNTHESIS OF INTEGRATED CIRCUITS (continued from page 17) Fig. 8 Focus regions (in the form of heatmap) of the network for the original and obfuscated image. On top of each predicted heatmaps, confidence score is mentioned. Fig. 7 Generated synthetic images of different cell types. Table 1 Overlap between layout and segmented synthetic SEM image Metrics Conditional Image Synthesis Mean IOU 0.94 Mean dice coefficient 0.97