Deep metric learning for sorting of image sequences– a solar cell use case

Alexandra Wörnhör

Thursday, November 07, 2019, 11:00

Room 01-012, Georges-Köhler-Allee 102, Freiburg 79110, Germany

Deep metric learning for sorting of image sequences– a solar cell use case

The idea of deep metric learning is to train a neural network such that feature representations are optimized to reflect given similarities of the input data. The similarities in the feature space should be easy to measure while a proper way in the input space is hard to find up to impossible. In this thesis, ideas of this approach are used and developed further to learn similarities between photoluminescence images of multi-crystalline Silicon (mc-Si) wafers. 

During the crystallization process of Silicon bricks defects develop continuously and affect the material quality. After slicing the Silicon bricks into wafers the defect signatures are visible in Photoluminescence (PL) images. The defect development can be used for process analysis and monitoring, as well as the design of experiments. Unfortunately, the correct wafer sequence cannot be tracked during production and is prone to errors due to wafer breakage and handling.

This work uses deep metric learning for the sorting of neighbouring wafer samples based on PL-images. In the first step, a meaning full representation of the samples is derived based on a triplet deep learning network architecture. Therefore, two novel loss functions are introduced: (1) a LAIT loss for ordinal data, which is used to optimize a network to detect the representation that belongs to the middle input image; and (2) RP loss for metric data designed to preserve the relations among the distances of the input data within a representation triple. 

In the second step, the representation is used for sorting sequences of wafers. Different approaches are used for networks trained with the LAIT-Loss and PR loss respectively.