IC Packaging Material Identification via a Hybrid Deep Learning Framework with CNN-Transformer Bidirectional Interaction.

Abstract

With the advancement of micro- and nanomanufacturing technologies, electronic components and chips are increasingly being miniaturized. To automatically identify their packaging materials for ensuring the reliability of ICs, a hybrid deep learning framework termed as CNN-transformer interaction (CTI) model is designed on IC packaging images in this paper, in which several cascaded CTI blocks are designed to bidirectionally capture local and global features from the IC packaging image. Each CTI block involves a CNN branch with two designed convolutional neural networks (CNNs) for CNN local features and a transformer branch with two transformers for transformer global features and transformer local-window features. A bidirectional interaction mechanism is designed to interactively transfer the features in channel and spatial dimensions between the CNNs and transformers. Experimental results indicate that the hybrid framework can recognize three types of IC packaging materials with a good performance of 96.16% F1-score and 97.92% accuracy, which is superior to some existing deep learning methods.