Infra-Red/UV Video Image
Segmentation Technique Theory
Mickael Maddison, May 2010
Currently the movie and photography industries utilize techniques often referred to as “chroma-key”, “luma-key” or “thermo-key” to film subjects for the purpose of removing the subject from the background of the image. Once the subject is removed from the background of the image, the subject can then be superimposed on alternative backgrounds. For example, filming an actor in front of a “green screen” and using the chroma-key technique to remove the actor from the green background would allow the video editor to place the actor on an image of the moon; without ever having to go to the moon.
Existing techniques for image segmentation require very careful and often expensive settings, lighting and filming techniques in addition to powerful post-production processing to achieve a quality result. This document proposes the use of the Near-infrared spectrum and optional ambient UV to replace the background; allowing for a much simpler means of extracting the desired image from an infrared background. Using a selection of isolated IR wavelengths and CCD or CMOS digital camera technologies adapted to capture and record these isolated wavelengths while at the same time recording the standard RGB (Red Green Blue) or RGBY (Red Green Blue Yellow) visible light would allow for software and devices to be produced will allow subject(s) to be removed from backgrounds with a higher degree of accuracy while requiring far less effort and processing.
* Adding in the detection of UV spectrum will also allow for additional processing options.
Uses: Cameras equipped with combined RGB/RGBY and IR/UV CMOS or CCD sensors would have a wide variety of uses.
1 – CCD: Existing CCD and CMOS sensors already have the capability to record near-infrared wavelengths. Most cameras use a special filter to block the infrared wavelengths from being captured. Cameras that do not have this filter in place store the infrared information in the RGB image. Currently, CCD and CMOS type sensors capture RGB light by using a special “Bayer Color Filter” as seen here:
Each square represents one “pixel” of information captured by the sensor. Processing techniques may vary, but in effect a square of 4 pixels are combined to create a single pixel of “true” color.
The following is an example of how a new filter could be designed to allow for the capture of the additional non-visible spectrum using the existing sensors:
In this sample image, instead of using a pattern of 4 pixels, the pattern is spread over 9 pixels. The 4 existing RGB pixels are captured in addition to 5 additional pixels as represented by the white and various shades of grey boxes. The optimal configuration is subject to analysis, but for example it could be laid out like this:
Red = Red, Blue = Blue, Green1/2 = Green1/2
White = wide-spectrum UV
lightest grey = 840nm IR, second-lightest grey = 900nm IR
third-lightest grey = 950nm, darkest grey = 1000nm OR wide-spectrum IR.
As a future consideration, there are also technologies coming that could utilize the optical properties of carbon nanotubes to capture information on specific wavelengths. Research has shown that a single carbon nanotube connected to a pair of electrodes can can measure IR radiation effectively. This technology may be a long way from practical use, however, it provides an ongoing opportunity to continue developing and refining the technology.
2 – File Format: In addition to the filter, a new image/video file format would be created to store the captured information in a useful format. Many cameras have built-in processors that convert the RAW pixel data from the CCD into consumer file formats such as mpeg, jpeg, tiff, etc. A new processor may be developed to provide traditional file formats + a masking file, or for more advanced use the camera may save all the data together to allow for more advanced processing and usage of the recorded data.
The new image/video RAW format would have more information available and would need to store unique information for each displayed pixel to be effective for image external processing.
3 - Software: Image and video editing software would require modifications and/or filters to be developed that would make full use of the extended information available through the new file format and/or the processing of the masking file with the video file. This may be modifications and additions to versions of existing industry-standard software. In addition, research may deem entirely new software should be developed to make full and wider ranging use of the data.
Some image processing systems have experimented using non-visible light to increase the accuracy and quality of the visibly produced image. For the purpose of image segmentation, the various wavelengths of non-visible light would be used to create highly detailed “mask(s)” useful for cropping the background from the desired image(s).
4 - Non-Visible Illumination: To make the most of the technology, a wide range of electronic devices would be developed as the technology is adopted by the relevant industries. Some examples of devices that would be developed and produced:
Benefits of the Technology
Patent and related technology research:
Live Action Compositing example http://www.scribd.com/doc/654481/Live-Action-Compositing
Using human-IR heat for image segmentation http://nae-lab.org/project/thermo-key/
This is not for IR - it is a device that uses a mapped background
Practical Example of Theory
Sony DCR DVD203 Digital Camcorder used in photograph mode.
IR narrow-beam floodlight (wide beam would be much more effective)
Blue floor mat for background
Doll with hair
A Dark room
Computer system with Adobe Photoshop Elements 8.0
Step 1 - photograph still image of subject in Nightshot plus mode which uses NIR to enhance image visibility.
Step 2 - photograph still image of subject in normal RGB mode (sorry for poor quality photo).
Step 3 - Open these to images into a single layered file in Photoshop (elements)
Step 4 - On the nightshot (Mask) layer, convert to B/W and increase contrast and in this case, with a blue background used, I adjust the blue level to get the best differential I can considering the poor lighting. You can see in this step that the segmentation opportunity is pretty good but not nearly perfect. The beam from the IR spotlight is a little too intense and focused. Developing proper Infra-Red lighting would not be difficult, at least in this scene.
Step 5 - Due to inadequate lighting I will cut excess dark regions (which would normally appear as a fairly even white background with proper IR lighting) and delete to pure white to match the area around the subject. As part of this step I have increased the contrast to show clearly the mask that was created. Proper lighting and some minor filtering would remove the manual portions of this step making it easy to automate.
Step 6 - Using magic wand (which would be an automated part of the process) I select all the white area and invert the selection to have the mask selected.
Step 7 - switch to the layer with the RGB image and copy the subject and paste into a new layer.
Step 8 - Turn off the visibility on all layers except the cutout of the subject. You now have a fairly good cutout of the subject to work with.
Step 9 - Insert background and adjust cutout layer as desired.
Due to nightshot plus mode captures IR data and incorporates this into an RGB image. Due to this, the Mask layer is not nearly as good quality as if an actual IR layer were saved in addition to the RGB layer. This limitation also requires the Mask layer of the subject to be shot in (visible) darkness to generate a strong contrast. The RGB layer is then shot in RGB mode with visible lights enabled. If I had 2 nightshot cameras of the same type, split the image into the 2 cameras, and had a filter on the RGB mode camera to remove all IR data and a filter on the nightshot mode camera to remove all RGB data I believe I would get a much more accurate mask.
For example, in unedited images you can see that the hair sticking up off the head is visible. With even IR lighting and no conflicting RGB data being stored on the mask layer, this should result in an even sharper, more accurate mask layer. The sharper, more accurate mask layer could then be used to cut out these fine details from the RGB subject layer, providing a nice crisp image to work with.
Also of note based on previous experience doing still-frame image compositing the process outlined in this document was very quick and simple to do. With proper camera(s), proper lighting, suitable software, and a good studio environment the results should be far better than this simple test.
|<< <||> >>|