Conclusion

Two demosaic algorithms, Alternating Projections and Primary-Consistent Soft Decision are examined in the project. From the experimental results, it was found that both of them achieved their goal. Ie, Alternating Projections managed to remove aliasing between the red and blue channels, and PCSD manged to demosic edges in a consistent direction. From the set of test images used, it can be seen that these artifacts are very common, and the ability to remove them definitely make these two demosaic algorithms worth implementing.

However, none of them is an all-time winner in the comparisons. Each of them perform better in some images and worse in the others. This shows that first, there is still room for improvement for both of these algorithms. Second, when we attempt to design a novel demosaic algorithm, or simply choose to implement an existing demosaic algorithm for a digital camera, most likely, there will be a tradeoff in what artifacts to remove.

Looking for subjective ways to compare the two different algorithms was not an easy task. Visual comparsion has always been the most popular way to compare the performance of demosaic algorithms. However, the problem with this is the result on visual comparison largely depends on the image used. To be more objective, a large number of test images of different properties can be used. Other comparison measures such as the MTF plot can also be used.

Future Works

It has been suggested during the presentation that trying to implement a hybrid of the Alternating Projections and PCSD might result in a demosaic algorithm with better performance since it seems that these two algorithms can help compenstate for each others' weaknesses.

More work can also be done on finding good ways to compare demosaic algorithms. This would likely help when one has to choose one particular algorithm to impelment in a digital camera.

Finally, currently, how images are stored in digital cameras is that a complete RGB image is first created through demosaicking. Then it is compressed to reduce the storage space needed. Recently, it has been suggested that a new approach to store images is to compress the mosaic image, and store the compressed mosaic image instead of the compressed demosaicked image [7,8,9]. Demosaicking can be performed when the user tries to retrive the stored image. Using this approach, demosaicking can actually be done outside the camera. More sophisticated demosaic algorithms can also be used. It has been shown that this approach reduces storage space, and give compressed images of excellent quality. When this approach is used, the robustness of the demosaic algorithms to compression becomes vital. Further work can be done to measure how compressed mosaic images affect the peroformance of the demosaic algorithms.

Acknowledgment

I would like to thank Prof. Brian Wandell and Joyce Farrell for their help on the project.