edfas.org 5 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 cluster (Fig. 2b), or a few fragmented signals due to cluttering by the circuitry layout or background contrast (Fig. 2c). In some cases, further challenges may include the presence of noise that may be associated to the setup environment, or the detection of very weak thermal signals. To manage data variability and noise removal, the algorithm implements the following four main steps: (1) detection of colored pixels from the gray background using Euclidean distance, (2) small particulate noise removal at pixel and connected component level, (3) clustering of nearby connected objects using density-based spatial clustering of applications with noise (DBSCAN),[4] and (4) identification of significant hotspots through ranking by size threshold based on the information from each image. The results of the algorithm based on 103 images containing 116 hotspots are discussed. For 86 images (83.5%), the hotspots were correctly identified as the only hotspot in the images. For 11 images (10.7%), the correct hotspots were identified along with other spurious hotspots. For six images (5.8%), the correct hotspots were not identified. In terms of hotspot detection, the algorithm achieves a sensitivity of 94.8% and has a false discovery rate of 17.9% (Fig. 3). Fig. 2 Three examples of images where the hotspot appears as (a) a single connected spot, (b) a diffused cluster, and (c) a fragmented spot that is broken up by the dark background. (b) (a) (c) Fig. 3 Results based on 103 images containing 116 hotspots. For the 17 images with incorrect hotspots detected, it was found that the spurious hotspots were big in size and thus, they were not removed during the noise removal and ranking by size steps. These hotspots were either comparable in size to the actual hotspots, hence, were identified along with the actual hotspots in 11 outcomes (Fig. 4a) or were larger than the actual hotspots, causing them to be selected in preference over the actual hotspots during ranking in the remaining six outcomes (Fig. 4b). It was noticed that the spurious hotspots were located outside of the package or die region. Therefore, the algorithm could be enhanced by implementing human-assisted identification on the region of interest before the application. IMAGE CLASSIFICATION WITH SUPERVISED LEARNING FA analysts typically reference the historical LIT images from the database to look for similar images. To improve the efficiency of image searching and automate image retrieval and ranking, a supervised algorithm has been developed to display similar images to a new query LIT image (Fig. 5). There are three main steps involved in this process: (1) algorithm for image similarity, (2) algorithm for image classification, and (3) re-ranking of images incorporating results from both image similarity and image classification algorithms. To work with a small dataset having only a few hundred images, a transfer learning method is used. Image features are extracted from the images by applying the pre-trained VGG16 network.[5,6] Principal component analysis (PCA) is further performed to construct image signatures that are relevant to this dataset. The top 100 significant PCA
RkJQdWJsaXNoZXIy MTMyMzg5NA==