I started with the single scale implementation. Although this method worked fairly well for the jpg images, it was rather slow. I initially used a shifting window of [-15 15]. Then I generated the multi-scale implementation using an image pyramid. I specified a depth and shifting value for the pyramid in order to have more control over the alignment process. Although this image pyramid worked well with most of the images, it caused alignment problems in some images. I initially tried edge-alignment, however that seemed to slow down my process a lot. Then I used a Gaussian Pyramid, which improved my algorithm. For the jpg images shown below, a used a depth of 1 (single-scale) and a shifting window of [-15 15]. For the tif images, I used a depth of 5 and a shifting window of [-5 5]. The single-scale implementation produced the results in approximately 1 second, whereas the multi-scale took approximately 24 seconds.
Bells and Whistles:
Auto Cropping: I cropped the borders of the images before the alignment by a factor of 0.1, in order to avoid long detections due to the borders. After the alignment, I cropped the image according to the shifting values, to get rid of the overlapping borders due to ‘circshift’.
Auto Contrast: I used Matlab’s ‘adapthisteq’ method, which implements contrast-limited adaptive histogram equalization on grayscale images. This function was used in the aligned grayscale images, before concatenating them.
| Images | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 | 11 |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Displacement | [3 3] [4 5] |
[5 2] [9 1] |
[4 1] [9 -1] |
[6 6] [11 8] |
[2 3] [5 5] |
[6 1] [12 0] |
[2 2] [4 3] |
[2 0] [6 0] |
[1 -1] [13 -1] |
[1 1] [4 2] |
[6 2] [14 4] |
| Images | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | 9 | 10 |
|---|---|---|---|---|---|---|---|---|---|---|
| Displacement | [39 15] [91 33] |
[47 38] [107 55] |
[35 25] [51 38] |
[14 6] [49 14] |
[57 25] [125 33] |
[16 2] [42 4] |
[50 13] [113 19] |
[-16 16] [11 17] |
[43 32] [111 59] |
[48 29] [101 49] |