This assignment was to combine several images with a varying depth of field, (ie, area of focus) into one image with an extended deth of field.
First, the input images from the provided data set.
Now, the result of extending the depth of field:
As you can see, the depth of field is extended, but we have produced significant artifacts. I hope that eventually I can find a way to eliminate them, but at the moment I'm with out recourse.
For my own dataset, I took a picture of some of my stuffed animals, arranged in my living room. The pathogens are up front, along with one of my candles. A monkey two mysterious lump-like things are on the coffee table, and "My Pet Monster" is way in the back on the couch. I had to reduce the size of these images signficantly from that recorded by my digital camera if I wanted the poor lab computers to finish processing sometime this century. Here are my input images:
What is the result of the operation?
The depth of field is improved, but poor Hapatitis looks like he has glasses! Some of the distortion is caused by camear movement (I really must invest in a good tripod). The rest by flaws in my algorithm. What can I do to improve the results? I have some idea, but unfortunately no more time to try them out right now. I could:
- Find a better way of determining what parts of each image are infocus. I currently use the gradiance, thEn create a mask with only the 99.99th precentile of the pixels remaining ('most gradient', the rest are zeroed). Then for the pixels that remain in the mask, I dilate them into a 30 pixel diamater disk. To produce the final image, for each pixel I take that pixel from the image that has the maximum mask value on that pixel. If I can determine a better measure of 'focusedness', it should imrpove the results
- Do a better job of reducing noise. I'm currently using a median filter on the input images to reduce noise. This reducuces the artifacts considerably, but does not do a good enough job.
- Find a better way to grow the masks. If you recall from point one, once I get the initial pixels in my mask (anchor pixels) , I blindly dialate them into a disk. This is why my artifacts have that distinct octagon shape. It might be possible to start with the anchor pixels and grow the mask outward in some intelligent way. Although it seems like if I had a better way to measure focusedness in the first place, this might not be necessary.
Finally, here is the matlab code that generated these images:
max_depth_of_field.m This takes the images as input, and returns as the result its best attempt at maximizing focus.
o_grad.m This is the functon that produces the mask for each image.