Improved Edge Detection Algorithm for Blurred Alignment Marks in Hybrid Bonding

The process of Cu-Cu hybrid bonding requires high-precision alignment between the bonded objects, such as a die and a wafer. In the alignment process, the alignment marks on the die and wafer are captured, and the positions of the alignment marks are detected from the image to calculate the misalignment. The previous method detected the position of the alignment marks by fitting a sigmoid function to the intensity profile around the edge of the alignment mark, and by increasing the size of the intensity profile, the amount of information was increased and the effect of image noise was reduced. However, because there are patterns around the alignment marks, there is a limit to how much the size of the intensity profile can be increased in order to achieve even higher accuracy. In this paper, we use multiple images to increase the amount of information and reduce the impact of image noise, thereby improving the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3\sigma$</tex> uncertainty of misalignment. Furthermore, by adaptively determining the initial values of the parameters for function fitting for each image, we can approach the optimal solution, and by reducing the number of times the parameters are searched for during fitting, we can reduce the computation time. In our experiment, even when the number of processed images was increased, the <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$3\sigma$</tex> uncertainty of misalignment was improved by 45% with the same computation time as the previous method.