Patterned Wafer Segmentation Pierrick Bourgeat ab , Fabrice Meriaudeau b , Kenneth W. Tobin a , Patrick Gorria b a Oak Ridge National Laboratory, P.O.Box 2008, Oak Ridge, TN 37831-6011, USA b Le2i Laboratory – Univ.of Burgundy, 12 rue de la fonderie, 71200 Le Creusot, France ABSTRACT This paper is an extension of our previous work on the image segmentation of electronic structures on patterned wafers to improve the defect detection process on optical inspection tools. Die-to-die wafer inspection is based upon the comparison of the same area on two neighborhood dies. The dissimilarities between the images are a result of defects in this area of one of the die. The noise level can vary from one structure to the other, within the same image. Therefore, segmentation is needed to create a mask and apply an optimal threshold in each region. Contrast variation on the texture can affect the response of the parameters used for the segmentation. This paper shows a method to anticipate these variations with a limited number of training samples, and modify the classifier accordingly to improve the segmentation results. Keywords: Wafer inspection, wavelet transform, segmentation, thresholding. 1. INTRODUCTION As semiconductor device density and wafer area continue to increase, faster and more sensitive automatic inspection tools are required. The size of the defects is becoming smaller, and harder to detect [1], [2]. This paper introduces an improvement of our previous work [3], on the image segmentation of electronic structures on patterned wafers to improve the defect detection process on optical inspection tools. Die-to-die wafer inspection is based upon the comparison of the same area on two neighborhood dies. The dissimilarities between the images are a result of defects in this area on one of the die. The two images are subtracted, and a threshold level is selected to locate any abnormality. This threshold is established upon the noise level in the difference image, to improve the signal-to-noise ratio. The noise level can vary from one structure to the other, within the same image since multiple structures coexist in the field of view. Therefore, the measure of noise within the whole image is not relevant for each individual type of structure. Segmentation is needed to create a mask of these different regions. This mask is then used to produce a measure of noise for each structure in the difference image, leading to an individual threshold for each region. For this work, segmentation is performed using the discrete wavelet transform [4] and the “à trous” algorithm [5], [6]. This algorithm is well adapted to discriminate local frequencies of the repetitive pattern, and it is restricted to principal directions that correspond to the geometric patterns found on integrated circuits. The weakness of this method is its sensitivity to contrast variation and small texture variation. In our previous work [3], a local correction was applied to remove the non-uniformities. This is sufficient in the case of small variations. However, in some cases where the variations become very important like large process variation or bad focus selection, the classifier needs to be trained with many different samples that cover all the variations contained within the die. The usual way to train a classifier on this type of data is an empirical approach. The classifier is trained with randomly selected samples, and then is tested over the whole set of data. The areas where the classifier performs poorly are used to extract new training samples. These new samples are added to the original set to retrain the classifier, until the best performances are obtained. This method is not realistic in the in-line inspection process when dealing with huge amount of data. It would require storing all the images of a die to process them off-line, and therefore a huge amount of memory would be needed. Meanwhile, it is time consuming for an operator to go through the iterative cycle of training the classifier, and testing its performances until they become acceptable. This paper introduces an original method to