edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 3 8 Layer 1 classifier consists of 298 images for training two classes whereas Layer 2 classifiers consist of 175 images for training five package classes and 122 images for training 16 die classes. Table 2 shows the performance of the algorithm on the images in the database (training set) and the query images (testing set). The algorithm is reviewed with and without the classification step to understand whether the twolayer classifier is beneficial in providing relevant images. From the observation at Layer 1, precision above 90% is achieved for both the training and testing set without using classification. This implies that the image similarity algorithm alone can filter images at the same package or die level as the query image. Classification improves the precision slightly. At Layer 2 Package, precisions of 64.8% and 72.7% are obtained for the training and testing set respectively without using classification. At Layer 2 Die, precisions of 41.7% and 45.6% are obtained for the training and testing set respectively without using classification. For this Layer 2 Package and Die, the classification delivers significant improvement in the performances (e.g., the training set (package) has 16.1% improvement and the training set (die) has 24.3% improvement). PROGRESS ON SIGNAL-TO-NOISE RATIO Further progress can be made to improve the signal-tonoise ratio by reducing the noise at the raw LIT image with software noise reduction. Different methods to suppress noise to improve the quality of the images are studied. These include characterizing the amplitude and phase of raw images available from LIT and the effects of exposure time on the signal-to-noise ratio (Fig. 8). With the suppression for noise, hotspots detection using the proposed algorithm can be accomplished at a shorter exposure time. CONCLUSION The image pattern recognition algorithm for detecting LIT hotspots not only benefits image processing, it can be leveraged to automate FA processes that require the identification of anomalies similar to hotspots in images as demonstrated in this work. Image classification with supervised learning has also shown favorable results. The image similarity search extracts image features and ranks them, while the image classification refines the ranking based on relevance for FA analysts. The precision of the algorithm is significantly improved after the classification process. To improve precision further, a larger dataset can be used. Fig. 8 (a) Original raw LIT image (b) after enhancement on image (a). (b) (a) Table 2 Precision for training and testing set with and without classification Precision, % Layer 1 Layer 2 Package Die Training precision without classification 92.0 64.8 41.7 Testing precision without classification 97.0 72.7 45.6 Training precision with classification 95.1 80.9 66.0 Testing precision with classification 97.3 76.2 54.3
RkJQdWJsaXNoZXIy MTMyMzg5NA==