My advice is to start with the simplest approach and run with it until you can't go any further. THEN look at the more complex approaches. You want to get results ASAP, without spending days scratching your head over insane code.
The local expert Image Analyst may have some real suggestions for you, but until he comes along I'll throw some simple ideas out...
The middle of those three pictures does not seem to be available, so maybe I'm confused because of that. Is the third image your background? Do you collect these from continuous video or do you just get a couple of snapshots?
Image 3 looks very different from Image 1. I know nothing about your field, but I'm going to assume that Image 3 is an initial reading, and Image 1 is a later shot after heating the concrete up.
First, you probably want to convert these images into greyscale if they're not already. I don't know what the formula would be for that heat pattern, but you just want one number for each pixel (maybe you have that already). To me it looks like it uses the entire blue range, then red comes in while blue goes out, then green comes in with the red maxed out.
If you want to try basic background subtraction, then normalise both images and take:
diffim = abs(image1-image3);
That'll probably be quite noisy so you would want a way to filter the noise out: ie detect minimum threshold. I'm no mathematician, but maybe standard deviation is useful here.
To do a global threshold, you just choose a number. Let's call it threshold. Then you cut all points above (or below) that. Since you have a difference image, you want to select only those with large differences:
defects = difim >= threshold;
Adaptive thresholding, as I understand it, is where your threshold value changes across the image. So instead of a single number, your threshold variable is another image. Presumably you'd want this because you need a larger difference at higher temperatures to consider a spot as a defect?
Anyway, you might generate this by convoluting the background image with a large-ish smoothing filter and taking the log of that, perhaps... That's the limit of my understanding of thermal energy! =)
smoothkernel = ones(20);
threshold = log(conv2(image3, smoothkernel));
Hope this is a help. It could be a start, if you currently have nothing =)
