Supervised Image Retrieval and Ranking Technique for Lock-in Thermography Images

Abstract

Lock-in Thermography (LIT) is a non-destructive technique in the failure analysis (FA) of integrated circuits (ICs). In diagnosing the cause of failure, a FA specialist spends a long time searching through a repository of historical images. In this paper, a supervised image retrieval and ranking algorithm incorporating image similarity and classification has been developed. Features are extracted from the images by passing them through the pre-trained VGG16 network. Principal component analysis (PCA) is then performed to identify 100 significant components that serve as signatures for each image and for computing Euclidean distance as the similarity metric. Next, a two-layer classifier replicating the human judgment process has been developed. The first layer of the classifier differentiates whether the query image is taken at the package or die level, whereas the second layer identifies the package or device class of the image. By analyzing the query image through the classifier, its classes in the two layers are determined. The distances of database images belonging to the same classes as the query image are reduced, shifting them ahead. The images thus sorted and ranked are recommended. The algorithm was tested on a dataset of 372 images of which 298 images were used for database construction, and 74 images were used as queried images. The incorporation of class classification improved the precision rate by recommending more images belonging to the same classes as the query.