I. Introduction
Defect detection is critical to the manufacturing of semiconductor wafers, yet relying on manual detection is time consuming, expensive and may cause yield ratio loss. A robust automated solution to this problem is essential, as the user will be shown only suspicious regions, thus saving valuable time. Defect detection is challenging as there are no precise characteristics of the possible defects and they may include particles, open lines, shorts between lines or other problems. Defects may belong to the wafer background or to its pattern, and may be predominant or scarcely noticeable. This variety makes it very difficult to perform template matching based on some a-priori features or training database of detects, and therefore encourages the development of unsupervised, data-driven methods.