An Active Deep Learning Method for the Detection of Defects in Power Semiconductors

Accurate detection of semiconductor defects is crucial to ensure reliability of operation and to improve yield by understanding and eradicating yield detractors. Recent advances in computer vision driven by Deep Convolutional Neural Networks (DCNN) and transfer learning have enabled novel techniques for defect detection and classification [1–7]. However, training neural networks requires very large datasets, even with transfer learning. This paper addresses this shortcoming by introducing for the first time the active learning approach for semiconductor devices. The proposed neural network can accurately identify defective dies with modest efforts in terms of annotating the image set. Finally, the feature maps of the DCNN are used to generate an unsupervised taxonomy of the semiconductor die defects, supporting further investigations to address yield detractors