edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 28 NO. 3 20 (a) (b) this methodology relies on the network’s ability to learn hierarchical features—from simple edges and textures in the early layers to complex hotspot shapes in deeper layers. By analyzing the pixel intensity distributions and local gradients, the model learns to distinguish genuine emission signatures from sensor noise and background artifacts. Representative detections are shown in Fig. 3 for both raw and overlay streams. QUANTITATIVE TWO-IMAGE COMPARISON A common FA task is distinguishing good versus bad units through systematic comparison. This task is automated through a quantitative two-image comparison pipeline designed to accommodate spatial misalignment, intensity variability, and noise. The workflow begins with image registration to establish spatial correspondence between reference and target images, followed by photometric normalization (e.g., background subtraction and controlled contrast/gamma adjustments). Comparison is performed at the object level rather than the pixel level to increase robustness to small residual misalignments. Hotspots are detected independently in both images, then matched using a one-to-one assignment strategy based on overlap (IoU) and centroid distance. For matched hotspot pairs, change metrics— including intensity difference, centroid displacement, and area or shape variation—are computed and aggregated into a composite abnormality score. Detections without a corresponding match are classified as “extra” hotspots (present only in the target image) or “missing” hotspots (present only in the reference image). The comparison categories are illustrated in Fig. 4. LAYOUT INTERPRETATION AND GDS-AVALON INTEGRATION To bridge the gap between image analysis and design context, a comprehensive layout interpretation system is employed to convert pixel coordinates into CAD coordinates and to integrate with the Avalon layout database for automated instance retrieval. The coordinate transformation process begins with extracting alignment metadata from setup data package files for Meridian Sierra format, which contain transformation matrices describing the relationship between physical devices and imaging coordinates. For overlay images, stored transformation parameters are used directly. For raw images, the pixel‑to‑CAD transformation is derived from available imaging and alignment metadata; when such metadata is unavailable, feature‑based registration is performed, followed by a quality‑gating step prior to exporting CAD coordinates. The Avalon integration establishes secure SSH tunnels with multifactor authentication (username/password and DUO 2FA) to access the hierarchical layout database. The system maintains local caches of frequently accessed data and performs periodic database scraping to stay synchronized with evolving design databases. Spatial queries are constructed using transformed coordinates in the format which is for further retrieving instance information including cell names, hierarchical paths, and electrical Fig. 3 Detected emission signals with bounding boxes from (a) raw grayscale PEM image and (b) overlay RGB PEM image. Confidence scores are generated by the model but omitted from the figure for visual clarity.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==