VRL an eye

Seamless Image Mosaicking


Dmitry V. Fedorov, Marco Zuliani, B.S. Manjunath


In order to successfully create visually pleasant image mosaics for aerial sequences or multi-temporal satellite images advanced mosaicking techniques are needed. Aerial photography sequences usually represent small miss registration due to transformation modeling which introduces effect like ghosting for mosaicking using blending or broken roads and abrupt color changes for normal mosaicking. In the case of multi-temporal satellite imagery mosaicking usual problem are abrupt color changes due to temporal changes like different crop stages on the field or object dislocation, etc. To successfully solve this problem automatically a two step procedure is used. In the first step we encounter the least error blending path among images over the overlapping area in order to generate seamless geometric blending. In the second step the seamless radiometric blending is achieved.


» On-Line Demo (include 3 ready to use examples)


Step 1: Identify least error path

In order to encounter seamless geometric blending we define the least error blending path among images over the overlapping area as the path that avoids high error regions as well as very uniform areas where radiometric differences are more visible. Ultimately the path should accompany object borders. Graph cut technique was used to solve this problem.

Img. 1a) Differences over overlapping area, image 1 in red and image 2 in cyan. Img. 1b) Regular blending creates object "ghosts". Img. 1c) Application of graph-cut to determine least error path, it can be visible on the right speaker.

Img. 2) Mosaic of two aerial sequence images with visible radiometric difference, it's possible to see how path follows object boundaries.

Img. 3a) Strong red on this difference image is mostly due to normal object movement caused by wind and in very small part due to transformation modeling error. Img. 3b) The observation here is that the border is better hidden in high gradient areas and is obvious in the uniform area. To overcome this problem the next step is needed.

Step 2: Render mosaic using multi-resolution spline

Once least error path is obtained we still have visible radiometric differences especially at the uniform areas. To overcome this problem multi-resolution spline technique is used. This technique splices images in different resolution levels preserving frequency components in their respective bands.

Img. 4) Multi-resolution spline of the same mosaic form Img. 2.

Img. 5a) The application of least error path. Img. 5b) Application of least error path followed by multi-resolution spline without any other disturbance to original images (i.e. no equalization, etc.).


Application to aerial photography, mosaic of the sequence of 7 frames:

Img. 7a) Mosaic performed without equalizing images.

Img. 7b) Zoom over urban area, more susceptible to miss registration.

Application to bio-imaging, mosaic of eye fundus images:

Img. 8a) Simple mosaic Img. 8b) Seamless mosaic of eye fundus


This project is supported in part by the Office of Naval Research, China Lake Naval Air Warfare Center and the NSF Information Technology Research grant #0331697.