Classifying Image Stacks of Specular Silicon Wafer Back Surface Regions: Performance Comparison of CNNs and SVMs

In this work, we compare the performance of convolutional neural networks and support vector machines for classifying image stacks of specular silicon wafer back surfaces. In these image stacks, we can identify structures typically originating from replicas of chip structures or from grinding artifacts such as comets or grinding grooves. However, defects like star cracks are also visible in those images. To classify these image stacks, we test and compare three different approaches. In the first approach, we train a convolutional neural net performing feature extraction and classification. In the second approach, we manually extract features of the images and use these features to train support vector machines. In the third approach, we skip the classification layers of the convolutional neural networks and use features extracted from different network layers to train support vector machines. Comparing these three approaches shows that all yield an accuracy value above 90%. With a quadratic support vector machine trained on features extracted from a convolutional network layer we achieve the best compromise between precision and recall rate of the class star crack with 99.3% and 98.6%, respectively.