Due to multi-lens camera utilizes each lens to record separate heterogeneous images, such as visible (RGB), multispectral (MS), or thermal (TIR) information, the differences of viewpoints and lens distortion leading to significant ghost effects of original images. To recover one-sensor geometry for precise spectral measurements, it requires band co-registration processing to correct the misregistration errors (MREs).
Based on our years research, we have developed a generalized tool, named Robust and Adaptive Band-to-Band Image Transform (RABBIT), that can accuratly correct the MREs of various state-of-the-art multi-lens cameras.
MREs of Original Image | Results of RABBIT |
---|---|
RABBIT supports various state-of-the-art multi-lens structured cameras as listed below.
Manufacture | Tetracam | Micasense Rededge | Parrot | |
---|---|---|---|---|
Camera | ||||
Models | MCA-4, MCA-6, MCA-12 | MX, Altum | Sequoia | Duo Pro R |
Multispectral | BLU, GRE, RED, REG, NIR | BLU, GRE, RED, REG, NIR | GRE, RED, REG, NIR | RGB + TIR |
*Please contact us if you need a different model.
RABBIT utilizes a novel N-SURF matching that can extract more features and increase the amount of correct matches on heterogenous images. This is a crucial step to connect images and estimnate the image transformation coefficients.
RABBIT adopts an extended projective transform (EPT) that can correct the differences of viewpoints and lens distortion effects. Compared to affine transform (AT) and projective transform (PT), EPT has the smallest residuals of MREs and best accuracy of co-registration results.
AT: Only parallel image rotation and translation are considered.
PT: Considering three dimensions axises rotation and translation.
EPT: Simliar to PT, but the lens distortion differences are also considered.
AT | PT | EPT |
---|---|---|
RABBIT can achieve 0.2-0.7 pixels band co-registration accuracy.
MCA-12 | Altum | Sequoia |
---|---|---|
RABBIT can run under multi-thread CPU or GPU. The comparisons of processing efficiency are listed in the following table.
CPU: Intel(R) Core(TM) i7-8700 3.2 GHz
GPU: NVIDIA Geforce GTX TITAN X
Camera | Tetracam MCA-12 | Micasense Altum | Parrot Sequoia |
---|---|---|---|
Image Resolution | 1280 X 1024 | 2064 X 1544 | 1280 X 960 |
Bands X Groups | 12 X 100 | 5 X 100 | 4 X 100 |
Independent in Multi-thread CPU | 65 min. | 62 min. | 17 min. |
Independent in GPU | 13 min. | 14 min. | 3 min. |
Batch in Multi-thread CPU | 2 min. | 2 min. | 2 min. |
Batch in GPU | 2 min. | 2 min. | 2 min. |
Camera | MCA-12 | Altum | Sequoia |
---|---|---|---|
Original | |||
RABBIT |
RABBIT is easy to use within three simple steps.
For example of Micasense RedEdge Altum: .../Images ....../Mica1 ....../Mica2 ....../Mica3 ....../Mica4 ....../Mica5
2.1 Read Camera Information (FLIR Duo Pro R Only) Since the sensors of FLIR has different image resolutions and focal lengths, it requires to import the interior orientation parameters in advance.
All images are loded and only master bands are shown on the left pannel. Click the Run buttom below and your results will sotre in the RABBIT folder.
Chose Camera Model | Select Image Patch | Run RABBIT |
---|---|---|
4.1 Feature Points The default value is 2% of the image resolution. Change to a higher or lower value to increase matching reliablity or reduce processing time. However, we do not recommend adjusting it, because a higher value of feature points will not siginficantly affect the processing time in GPU mode, but it does in CPU mode.
4.2 Independent vs. Batch Independent is to conduct N-SURF matching and EPT on each group of images, which can obtain more accurate and reliable results. Batch is to use same coefficients from one corrected group of images, which can achive ultra-speed processing efficiency.
When to use batch mode If all lenses of camera are highly synchronized, have same shutter speed, and the observation object distances are constant, we can treat the imaging geometry among all groups of images are the same. Therefore, we can simply choose one group of image for correction and then apply the obtained coefficients to the rest groups of images.
** Tips: Choose one group of images that have rich texture and high contrast for correction.
Setting | Accuracy of one group |
---|---|
4.3 CPU vs. GPU By default, RABBIT detects your most poweful GPU device for increasing the processing efficincy.
We use Alea GPU commuity for GPU accelerating, see what kinds of GPU cards are supported in Alea GPU
If RABBIT help in your research/projects, please consider citing the following papers.
If you have any questions about RABBIT, please contact Dr. Jyun-Ping Jhan jyunpingjhan@geomatics.ncku.edu.tw