3.4 Experimental Setup 35 Figure 3.2: The full 3D-printed product. We are only interested in defects on the top surface (i.e. the outer ring), so grey-scale images of the full product, obtained from a flatbed scanner, were chopped up into smaller images of 224x224 pixels each, centred and rotated to show only a part of the surface. Each of these smaller images was then manually labelled as either acceptable or containing some defect. We distinguish five classes of defects, as illustrated in Figure 3.3. Preprocessing also included gamma correction, to make potential defects more visible. Table 3.1 shows the number of images in each defect class. We use the first 500 non-faulty images as a training set, and the remaining 119 as a test set. We use all the images from the defect classes as anomaly test sets. Table 3.1: Number of data points per defect type for the 3D-printed products dataset. Class: Good Shrink hole Dirty spot Discoloured line Edge Erosion Print line Quantity: 619 138 75 24 239 190
RkJQdWJsaXNoZXIy MjY0ODMw