3.5 Results 45 (a) Shrink hole. (b) Dirty spot. (c) Discoloured line. (d) Printed line. Figure 3.11: ELBO density distributions for normal (blue) and anomalous (red) 3Dprinted products with a 64-dimensional latent space. decide whether an anomaly is acceptable or not, in particular in doubtful cases (i.e. cases close to the decision threshold). The examples in Figure 3.12 also give an indication why certain anomaly types appear easier to detect than others, using our method. In particular, defects shrink hole, edge erosion, andprint line, which were detected best according to Tables 3.5 and 3.6, appear to affect more pixels than the other two defects. 3.5.3 NLST 3D nodules GAN results Figure 3.13 shows examples of new data produced by the GAN. While comparing different generated images from the variations in training of the 3D WGAN-GP, we notice a partial mode collapse (Goodfellow, 2016) in the samples. We can
RkJQdWJsaXNoZXIy MjY0ODMw