It is frequently desirable to combine different sources of image information into a composite image prior to undertaking image analysis. For example, multiple images may be merged to extend the field of view or resolution, or to eliminate foreground obstacles. Or stereo images may be combined so that regions occluded in one camera's view are filled smoothly with regions seen by the other camera. The essential problem in image merging is "pattern conservation": important details of the component images must be preserved in the composite while no spurious pattern elements are introduced by the merging process. Simple approaches to image merging often create edge artifacts between regions taken from different source images, and these may confound subsequent image analysis. We describe an approach to image merging based on pattern decomposition. Each source image is first transformed into a set of primitive pattern elements. Pattern sets for the various source images are then combined to form a single set for the composite image. Finally the composite is reconstructed from its set of primitives. We illustrate the pattern decomposition technique with several practical applications. These include image merging to eliminate foreground objects, and merging to extend the depth of field. In all cases the Laplacian pyramid is used to encode images in terms of sets of primitives which resemble Gaussians of many scales.