Hi,

I was trying to put together some code to determine wither two photographs are different i.e. someone is standing in one and not the other.  Write now this is what I have - it returns a percentage of colour difference and generally if it's above 50% there is something else in it.

This code is ridiculously slow to run and takes about a minute or so... I need it in seconds!  It's even slow with the thumbnail sized images.

I'm not looking for anything hugely accurate - the big problem doing this is clearly that a pixel is going to be different depending on light levels.

I'd be really grateful for any suggestions, feedback or an existing solution that I can easily use.

I'm not a coder - so I'm trying to keep it simple - feel free to tell me what im doing wrong!  It's good to share.

     void compareImages ()
        {
            Debug.Print("Comparing Images");

            int height = camera.CurrentPictureResolution.Height;
            int width = camera.CurrentPictureResolution.Width;
            int DiferentPixels = 0;
  
            Bitmap container = new Bitmap(width, height);

            for (int i = 0; i < width; i++)   // increment by ten in width
            {
                for (int j = 0; j < height; j++)  // increment in ten by height
                {
                    Debug.Print("doing...");
                    Color currentPixel1 = picture1.GetPixel(i, j);
                    int r1 = ColorUtility.GetRValue(currentPixel1);
                    int g1 = ColorUtility.GetGValue(currentPixel1);
                    int b1 = ColorUtility.GetRValue(currentPixel1);

                    Color currentPixel2 = picture2.GetPixel(i, j);
                    int r2 = ColorUtility.GetRValue(currentPixel2);
                    int g2 = ColorUtility.GetGValue(currentPixel2);
                    int b2 = ColorUtility.GetRValue(currentPixel2);

                    if (r1 != r2 && g1 != g2 && b1 != b2)
                    {
                        DiferentPixels++;
                        container.SetPixel(i, j, GT.Color.Red);
                    }
                    else
                        container.SetPixel(i, j, picture1.GetPixel(i, j));
                }
            }
            int TotalPixels = width * height;
            float difference = (float)((float)DiferentPixels / (float)TotalPixels);
            float percentage = difference * 100;

            oledDisplay.SimpleGraphics.DisplayImage(container, 0, 0);

            Debug.Print("Difference: " + difference + "Percentage Difference: " + percentage);

        }

I think your right.

I was looking to detect movement in an image in order to know were people are around 360 degrees.

I've been told the best way is to upload the image and have a web server do the processing - here I could use existing API's.

My attempt was for a quick and easy solution!

For starters, take out Debug.Print statements from your code. Those are INSANELY slow. Useful when trying to figure out what your code is doing, but super slow. Next thing, make sure not to run your code in the debugger. Thirdly, you're constructing a new image with "changed" pixels for display--that is another huge drain on your processing (the display part primarily).

As far as the algorithm itself... If you truly want to compare colour information between the two pixels, then just compare the colour values. Don't spend 6 function calls decomposing the individual colour channels, and then 3 operations comparing them for equality. The whole thing can be done in a single operation: colour1 == colour2. Done.

Finally, there is no simple way to do image processing. It has been the bane of computer scientists for a long time. With the new-found popularity of the GPGPUs massively parallel computations are making the pain a bit easier, but in a small CPU as you have, there is very little you can do, but wait for your code to process pixels one at the time.

If connecting to the Internet is an option for your project, you could offload the image processing to a Web service.  It's higher-latency, but it saves a lot of local compute power.  Also, a Web service can store more images and look for differences over time. 

There are some good tutorials on connecting Gadgeteer devices to Web service at in the Featured Projects area.  The Text-to-Speech Translator is a great model (though it uses a pre-existing Web service and you'd probably need to build something custom).

Thanks! Yeah, made it a while ago.

 

P.S. any idea why I can't set my avatar on tinyCLR forums?

I did something similar myself but concluded Gadgeteer wasn't up to the job (I was porting code that worked fine in native C on somewhat similar-ish hardware). My method was to calculate average colour values for 10x10 pixel blocks (in order to filter out the noise and jitter), and store reference arrays of those averaged block colours. To compare one photo against a reference, perform the calculations on the pixels to obtain a new set of averaged colours, then work out abs(old.R-new.R)+abs(old.G-new.G)+abs(old.B-new.B) for each block position, if it's above some threshold you can count that block as 'different', and tot up how many blocks were 'different' to decide if you're detecting movement. Replace your reference image every so often in order to cope with changes in brightness as the time of day changes and the sun moves around. Unfortunately though, code that runs fast enough in C to process a 320x240 camera frame buffer many times a second, takes minutes per frame in C# in Gadgeteer. If the hardware is Fez Spider, maybe it's possible to do natively in RLP?