Finding a specific edge in multiple images?

I have a set of fotos taken by a camera that shifts slightly in the wind. I am trying to correct the shift automatically. This is the image I want to use as reference: http://i101.photobucket.com/albums/m80/klar_01/marks.jpg

This is an example of a shifted image: http://i101.photobucket.com/albums/m80/klar_01/shift2.jpg

I have cropped the reference image and isolated a part of the main horizon line using canny edge detection and a threshold. I would like to find this edge in the other images and then transform the images to line up with the reference. Unfortunately the lighting situation changes a lot and the edge can not always be picked out clearly.

This is the horizon line from the reference image (green) and the canny edges I get for the shifted image (purple). http://i101.photobucket.com/albums/m80/klar_01/edges.jpg

I can see where the green and pruple lines should line up but I have no idea how to express this in code, as I can't get a clear edge in the shifted image.

Any suggestions would be much appreciated. Other ideas about how to approach the problem would also be very welcome!

I have the computer vision toolbox and have tried using the inbuilt feature matching functions with SURF features and corner points. Points do not match up correctly, presumeably because of the changing light.

[Lea]

anyone?

Answer 1

[Image Analyst]

Did you try imregister()?

[Lea]

Thank you! The following code seems to produce acceptable results for the uncropped images linked in my question.

    K=imread('marks.jpg');
    L=imread('shift2.jpg');
    movingg=rgb2gray(K);
    fixedg=rgb2gray(L);

optimizer = registration.optimizer.RegularStepGradientDescent;

metric = registration.metric.MattesMutualInformation;

movingRegistered = imregister(movingg, fixedg, 'rigid', optimizer, metric);

I have 1000s of images and would like to speed up the process, so I was thinking I could use cropped images that show only a part of the horizon, obtain the transform with imregtform(), then apply that to the uncropped image. Like so:

    rect=[1069.5 4.5 979 333];
    fixed=imcrop(fixedg, rect);
    moving=imcrop(movingg, rect);

    optimizer = registration.optimizer.RegularStepGradientDescent;

    metric = registration.metric.MattesMutualInformation;

    tform = imregtform(moving, fixed, 'rigid', optimizer, metric); 

    movingRegistered = imwarp(Lg,tform,'OutputView',imref2d(size(fixed)));

This does not produce an error message but the registered images are more missaligned than they were to begin with. I can not get the cropped images to register correctly, let alone register the full image using the transform from the crops. I am guessing I am not understanding something about how the metric works? Is there some other flaw in my thinking?

[Image Analyst]

Sorry - I haven't used the imregister function as much as you have, so I can't help any more. You know more about it than me at this point.

[Harshan Desu]

Hello Image Analyst, can u suggest me a way find and extract only a specfic edge from one image. Please help me out.

https://www.mathworks.com/matlabcentral/answers/523335-how-to-extract-specific-edges-from-an-image?s_tid=prof_contriblnk

Answer 2

[Alex Taylor]

Lea, I will look into this a bit, I'm busy with other things at the moment. I do have a couple other questions for you though:

1) Did you try the 'multimodal' configuration from imregconfig? Because MutualInformation is a noisy metric, pairing MutualInformation with RegularStepGradientDescent can be unstable from for some input images. Sometimes it works fine. A more typical configuration is to use the OnePlusOneEvolutionary optimizer with MutualInformation metric.

2) Am I correct that you are solving 2-D problems only? This is not volumetric registration?

3) Are you solving 1000s of independent 2-D registration problems, or does the transformation estimate from one set of images in your dataset tell you anything about the transformation you will expect for subsequent registration problems?

I'll try to mess around with your images when I get a minute and see if I can come up with anything better. In the meantime, you might take a look at the Image Registration GUI from the File Exchange. It is useful for sweeping through the parameters of the metric/optimizer configurations.

http://www.mathworks.us/matlabcentral/fileexchange/34510-image-registration-gui

- Alex

[Lea]

Alex, thanks!

1: I have tried all sorts of combinations and RegularStep with MutualInformation seems to work best. I have no idea why as this is pure trial and error, so I am very grateful for your explanation and will look into multimodal configuration some more.

2: Yes, 2D only.

3: The transformations will all be similar but not exactly the same and they are essentially independent, yes.

Maybe some more explanation: I have a year's worth of images from an automatic camera that takes a picture every two hours. In strong winds the camera wobbles a bit so images may or may not be shifted depending on the undefinable camera wobblyness.

I need to automatically detect snow in certain spots on the pictured slope. I would like to define the relevant ROI once and just check this position for every image (that seems easiest). The ROI is only a few pixels large so even a small image missalignment will mess up results. (The yellow things in the linked reference image mark the spots of interest)

I have gotten the cropped images to register with the above configurations and setting

metric.NumberOfHistogramBins= (10)

but have not manged to successfully apply the transform from that to the full image.

Any suggestions about how to register the images in a time effective manner would be hugely appreciated. I will check out the GUI now. Thanks again.

edit: I just checked and I don't actually have thousands of pictures, just about 1500, thanks to darkness and fog...

[Alex Taylor]

Lea,

To my eyes, it looks like I can get reasonable registration results by simply accounting for translation in the full sized images. Do you have reason to believe there is also rotation in your camera due to wind? It's hard to tell from only having the two images. You mention "shift" in the camera, which I interpret to mean translation, but I see you are using a 'rigid' model in imregtform...

[Lea]

Alex, thanks so much for your input.

It's probably just translation, you're right. I'm not 100% sure as I didn't set up the camera and don't know how it can move, so I guess I was being extra careful. changing that would be an easy way to make it quicker, I'll investigate the camera rig.