edfas.org 7 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 two classes as the query image are reduced by a fixed constant. Results from both layers of classifiers are used to rank higher to differentiate with those from a different class. Finally, the images are re-ranked and recommended to the FA analyst. The algorithm is tested on a dataset of 372 images that is divided into two subsets. 80% of dataset (298 images) are used to construct the database and the other 20% of dataset (74 images) are used as query images to evaluate the training and testing accuracy of the algorithm. Among (a) the 298 images, 176 images and 122 images belong to the package and die level, respectively. For the evaluation, the precision “k” is fixed at five, in other words, the top five recommendations from the number of relevant search results are selected. Table 1 shows the optimal number of PCA components and the corresponding cross-validation accuracies for the two-layer classifier. By selecting 10 PCA components as the first layer classifier, the highest accuracy attained is 92.2%. For Layer 2 package classifier, 20 PCA components are required to achieve an accuracy of 77.6%. For Layer 2 die classifier, 50 PCA components are required to achieve an accuracy of 47.5%. More PCA components are required to differentiate the finer features in each image. Higher accuracy is attained with larger number of training images and smaller number of classes. For instance, (b) (c) (d) (e) Fig. 7 Layer 2 package classifier differentiates the five package types (DIP, DSO, VQFN, LQFP, and BGA). Table 1 Cross-validation accuracy for Layer 2 classifier Classifier Optimal number of PCA components Cross-validation accuracy Layer 1 10 92.2% Layer 2 package 20 77.6% Layer 2 die 50 47.5%
RkJQdWJsaXNoZXIy MTMyMzg5NA==