This paper presents a novel technique for detecting defects in periodic 2-D wafer images when there is no image database or priori knowledge. It creates golden block database from the wafer image itself and customizes its content when needed.Spectral estimation is used in the first step to derive the periods of repeated patterns in both directions. Then a building block representing the structure of the patterns is extracted. After that, a new defect-free image is built based on this building block. Finally, a pixel-to-pixel comparison is all we need to find out possible defects.The extracted building block is stored as the golden block for certain pattern. When a new image with the same periodical pattern arrives, we don't have to re-calculate its periods and building block. They can be derived directly from the existing golden block. It is a bridge between the existing self-reference methods and image-to-image reference methods.