Defect detection and classification of silicon solar wafer featuring nir imaging and improved niblack segmentation
Loading...
Date
2016-04-01
Authors
Zeinab Mahdavipour
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Producing a high yield of renewable energy requires a high efficiency in
product fabrication of silicon wafers, which is the basic building component of solar panels. For this reason, the high quality inspection of solar wafers during the procedures of production is very important. In this thesis, an automatic and efficient defect detection system, utilising advanced classification and clustering strategies are proposed. In this study a machine vision scheme for detecting micro-cracks and other
defects in polycrystalline and monocrystalline solar wafer manufacturing is proposed and developed. Micro-crack inspection is very challenging, because this type of defect is very small and completely invisible to the naked eye. The presence of other heterogeneous structures in solar wafers like grainy materials and dark regions further complicates the problem. In this study an efficient micro-crack inspector comprising near infrared illumination and an improved Niblack segmentation algorithm is proposed. Empirical and visual results demonstrate that the proposed solutions are competitive when compared to existing Niblack thresholding formulas
and other standard methods, and achieve better precision and performance in terms of Pratt’s figure of merit and other evaluation methods. Result in a figure of merit (FOM), accuracy (ACC), dice similarity coefficient (DSC), and sensitivity were consistently higher than 0.871, 99.35 %, 99.68 %, and 99.75 %, respectively, for all images tested in this study. Meanwhile, a set of descriptors corresponding to Elliptic Fourier Features shape description is extracted for each defect and is evaluated for each cluster to use for clustering and classification part. The classification combines the analysis of defect intensity features, the application of unsupervised k-mean
clustering and multi-class SVM algorithms. The methods have been applied for detecting, clustering and classification polycrystalline solar wafer images, corresponding to defects such as micro cracks, stain, and fingerprints. Results indicate that the k-mean and SVM classifier can accurately cluster the defects with accuracy, Rand index, and Silhouette index averaging at 99.8 %, 99.788 %, and 98.43 %, respectively.