August_EDFA_Digital

edfas.org ELECTRONIC DEV ICE FA I LURE ANALYSIS | VOLUME 24 NO . 3 16 CELL IMAGE SYNTHESIS To address the data deficiency problem for the end- to-end system, cell images have been synthesized by the generative adversarial network (GAN) using cell images acquired in the previous section. In an adversarial set up, a generator G learns a distribution p g over data x by building amapping function from the prior noise distribution p z ( z ) to dataspace as G ( z ; θ g ). The discriminator D ( x ; θ d ) outputs a probability value representing the probability of x came fromtraining data rather than p g . G and D are both implemented using CNN in this case. G and D both are trained simultaneously. Throughout learning, D is trained to maxi- mize the output of assigning correct labels to both real (x) and synthetic images ( G ( z )) i.e., it maximizes the expectation log D ( X ). On the other hand, G is trained to fool the network by generating images closer to real images by minimizing log (1 ‒ D(G(z ))). So, it seems like both G and D get involved in a two-player minimax game with value V ( D , G ). (Eq 2) In this case, generated synthetic images are conditionedonclasses. So, the class label information was encoded with the sampled random variables and passed through the (a) Fig. 4 (a) C i represents row-decider components and R k represents a cell row (marked by two yellow lines). (b) Single and composite cell separation from SEM image row. Consecutive components are merged based on pixel level distance in steps (ii) and (iii). Single and composite cells are separated based on the newly merged components in steps (iv) and (v). (b)