Image Processing: Pixel Classification

Good day,

I have recently been tasked to a project which primarily deals with image processing. Given a sky image (clouds and blue sky etc.) I am supposed to identify which pixel is cloud and which is sky. I am absolutely new to the program and I would like to seek some professional advice from the community to point me in the right direction on what I should be looking at or doing.

Feel free to drop me any comments. Any help rendered is deeply appreciated.

Best Regards Guan Zhao

16/09/2012

Answer 1

[Image Analyst]

This can be done somewhat effectively with the color classification demos in my File Exchange: http://www.mathworks.com/matlabcentral/fileexchange/?term=authorid%3A31862. However, as you can imagine, as the clouds get thinner and thinner, it's really not clear what is cloud and what is sky. What about really thin wispy low density clouds, or haze, or fog? What is this project? Is it a simple student project? If so, you'll probably be supplied with high contrast dense white clouds and my code will work for you. If it's a real-world project for your employer (for example NOAA), then you'll have to do a lot of research. I saw a paper on could identification given this past January at the SPIE Electronic Imaging meeting in San Francisco. Look up papers like that. For example in section 16.6.2.11 Tracking for Weather, Clouds here: VisionBibliography

[Guan Zhao]

It's a high weightage student project, and yes I do get what you mean on thin clouds. Been reading up research papers though its very hard to comprehend what they are saying. A paper I read mentioned using probability measurements and neural networks to identify cloudy pixels but I feel its too difficult to understand and implement for this project.

The second paper mentioned about using a pixel's Red-Blue Ratio to determine if a pixel can be considered as cloudy, and I thought this technique might be useful to my project.

My project requires me to analyze a given ground-based sky image for cloud identification. So far the images provided to me shows very obvious cloud patterns and sky contrast. However, the contrast will differ according to the time (e.g. blue sky turning yellow in the evening). And yes, as the project progresses, I am sure the thin cloud challenge may come into context.

What's your take and advice for me? I am new to MatLab, so my concern for now is whether the learning process plus the implementation process will exceed my given time frame. Would deeply appreciate your advice!

Regards

Guan Zhao

[Image Analyst]

I'd recommend you use either the HSV color segmentation method or the Delta E method. Either should work.

You can also look at the color gamut inspection tool: http://rsb.info.nih.gov/ij/plugins/color-inspector.html You can run it from MATLAB or imageJ. It will let you visualize the gamut and that will help you figure out what part of color space your clouds lie in.

[Guan Zhao]

Will do. I'll post when I have questions. Thanks a lot.