edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 6 components are taken out from the 4096 features out- put from VGG16. By implementing transfer learning together with PCA, the model is optimized for the small set of images. A two-layer classifier, designed to replicate the decision-making process of the FA analyst, has been developed to further improve the ranking algorithm. Upon receiving a query image, the first layer of the classifier differentiates between whether the image relates to the package or die level (Fig. 6). The second layer of the classifier then categorizes the image under the appropriate package class (if it is captured at the package level) or die class (if it is captured at the die level). This dataset contains five package classes (Fig. 7) and 16 die classes. Next, the algorithm extracts relevant images captured at the same package/die level belonging to the same package/die class. The classified images are then ranked according to the Euclidean distance from the query image. The Euclidean distances of images that belong to the same Fig. 4 Two examples where the hotspot recognition did not work. (a) The hotspot was correctly identified (near the center region) along with noise (at the edge). (b) The wrong hotspots (in red bounding boxes) were identified. The actual hotspot (in yellow circle) was not identified by the algorithm. (b) (a) Fig. 5 The methodology. The outline of each image corresponds to its class identity. Fig. 6 Layer 1 classifier determines whether images are taken at package-level (top) or die-level (bottom).
RkJQdWJsaXNoZXIy MTMyMzg5NA==