0

I'm writing software for a solar panel inspection system and need to stitch camera images taken in a electroluminescence machine together. These images have only very few features and low contrast (as shown below) and the OpenCV image stitcher does not work by itself. left image right image The approximate overlapping area of both images is known but I need the result as accurate as possible. I tried shifting one image over the other and computed different distance measures over the ROI, but without satisfying results. The SSD distance does not work due to vignetting and differences in the pixel intensity. Normalized Gradients or Cross Correlation were not robust either.

Any idea how to preprocess the images for the stitcher to work? Or is there another way to tackle this? Dark cells are not always present, making them not reliable as a feature.

There are vertical cell edges (nearly invisible due to bad contrast), but I have no method of detecting them either. If these could be detected, I could align the images with the cell edges.

Any help is much appreciated.

ftramnitzke
  • 151
  • 9
  • you'll need at least 4 point correspondences without any further assumptions, to compute a perspective homography. These repeated structures are not so nice for automatic stitching ;) – Micka Aug 04 '17 at 11:29
  • I'm fine if it's not automated and I'll code it myself, but I need an idea on how to do that ;) – ftramnitzke Aug 04 '17 at 11:39
  • use those two dark areas to choose 4 or more 2d points, then compute perspective homography, then use cv::warpPerspective – Micka Aug 04 '17 at 11:41
  • This might work, but not all images have these dark cells as stated in the question. – ftramnitzke Aug 04 '17 at 13:11
  • add samples of images without dark cells then. The overlap is quite small, too. – Micka Aug 04 '17 at 13:13
  • Hey, have you found a solution how to stitch those kind of images smoothly? – gameon67 May 29 '19 at 09:07
  • Hey @gameon67, in the end I did this completely manual by firstly reducing the effects of vignetting through lens distortion correction. Afterwards I manually shifted the single camera images over one another until I had a visually satisfying result. This machine works automated so the positions of each camera image stays the same and once set up, the stitched results where acceptable enough. – ftramnitzke Jun 24 '19 at 07:19

0 Answers0