Correction of Chromatic Aberration from a Single Image Using Keypoints

02/08/2020
by   Benjamin T. Cecchetto, et al.
0

In this paper, we propose a method to correct for chromatic aberration in a single photograph. Our method replicates what a user would do in a photo editing program to account for this defect. We find matching keypoints in each colour channel then align them as a user would.

READ FULL TEXT VIEW PDF
POST COMMENT

Comments

There are no comments yet.

Authors

page 2

page 3

page 4

04/17/2017

Multimodal Prediction and Personalization of Photo Edits with Deep Generative Models

Professional-grade software applications are powerful but complicated-ex...
09/26/2020

Interactive White Balancing for Camera-Rendered Images

White balance (WB) is one of the first photo-finishing steps used to ren...
12/18/2009

Matching 2-D Ellipses to 3-D Circles with Application to Vehicle Pose Estimation

Finding the three-dimensional representation of all or a part of a scene...
09/12/2016

Generative Visual Manipulation on the Natural Image Manifold

Realistic image manipulation is challenging because it requires modifyin...
06/06/2020

Simple Primary Colour Editing for Consumer Product Images

We present a simple primary colour editing method for consumer product i...
11/19/2020

Multi-Plane Program Induction with 3D Box Priors

We consider two important aspects in understanding and editing images: m...
09/04/2019

Program-Guided Image Manipulators

Humans are capable of building holistic representations for images at va...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

1. Introduction

Chromatic aberration (also known as colour fringing) is a phenomena where different wavelengths of light refract through different parts of a lens system. Thus, the colour channels may not align as they reach the sensor/film. This is most notable in cheaper lenses, it is also noticeable at higher resolutions. We desire an image free of chromatic aberrations is simple, so all the planes are in focus. For a simple lens system this amounts to misaligned colour channels (red, green and blue). The misalignment is due to a uniform scaling and translation. There are many types of other chromatic aberrations. Complex lens assemblies introduce new distortions. We propose a method to deal with the simple, more common scenario. We later discuss possible ways to deal with more complex ones.

(a)
(b)
Figure 1. LABEL: A photo with chromatic aberration. LABEL: A correction for it using our algorithm. (Foundation, 2009)
(a)
(b)
Figure 2. values over the images in Figure 1.
Figure 3. Colour points ) in neighbourhood clusters around a pixel. Left: From a control image with no aberrations chosen randomly. Right: From a photo with aberrations, chosen at regions of obvious colour fringing.
(a)
(b)
Figure 4. Zoomed in area from Figure 1. LABEL: Original photo. LABEL: Corrected.

2. Previous Work

Many domains have sought to correct chromatic aberration. A major goal is to improve accuracy of measurement. This could be either in a lab or more constrained capture scenarios such as of space. One such method is in active optics (Willson and Shafer, 1991). For active optics, instead of taking one exposure for the colour channels simultaneously, the user takes one for each exposure at different times. Each time, the focus changes to match the channel’s optimal plane. This accounts for the aberration in the lens so the three channels will align in the final result. This requires a change from traditional hardware as well as three exposures at three points in time. This limits capture for dynamic scenes, as the object may be moving in that time span.

Another proposed method is to model the edge displacements with cubic splines to compute a nonlinear warp of the image (Boult and Wolberg, 1992). This method cannot handle blurring/defocus in the image plane along with saturation effects. It breaks down in regions that are either underexposed or overexposed. A different technique exists for correcting lateral chromatic aberration where one calculates the aberrations and uses warping to compensate (Mallon and Whelan, 2007). This only works for calibrated cameras/images which may not be workable for a more simple user.

Chromatic aberrations can also determine depth information without prior calibration (Garcia et al., 2000). Recently, an analysis of aberrations show they are not a simple shift of colour planes. We can get a corrected result by computing a non-linear warp (Kang, 2007). This algorithm handles many different categorized cases of aberrations. Its only limitations are artifacts caused by saturation, and slow, non-linear, performance.

Since we want to deal with the simple case, we may want to remove perspective distortions. A few algorithms have solved this problem in the past (Swaninathan et al., 2003). If we undistort a lens distortion, we should end up with only needing to solve for the translation/scaling of the colour channels to align them.

A recently proposed metric, , can determine a depth matte from a colour-filtered aperture (Bando et al., 2008). This metric measures how collinear and correlated points are in 3D (colour space). It is then used to find a disparity from their custom colour-filtered aperture.

3. Overview

Instead of concentrating on the optical derivation, we instead consider an image-based one. An artist can align colour channels of an image to correct the aberration. They align the channels by moving the corners of the images to transform each channel’s image. We attempt to automate this artist driven result.

The main idea is to align two of the colour channels to the third one. The green channel has the least amount of aberration as it is in the middle of the colour spectrum. In some of the results, we match to the red channel for simplicity. Choosing a different channel to match to results in a scaled version of the corrected image.

Our solution is to find keypoints in all channels, and find where the keypoints match to the fixed channel. At the same time, we minimize the aberration metric . This is like finding disparities in stereo, but in 3 channels instead of two images. Once we have these matchings, we can prune the results based on how well the matchings are. Then we need to find a transformation from the non-fixed colour planes to the fixed one. We restrict the transformation to a scale and translation. We solve the issues that arise from saturation by ignoring saturated regions and their neighbours. Since the algorithm is linear, it is quite efficient.

3.1. Computing the Alignment Metric

Before we start describing how to find the keypoints, and disparities, it is best to define how aligned the colour channels are. It is not suitable to do cross-correlation, as that only tells us how aligned two colour channels are. We want to see how aligned three are. As described above, there exists such a metric  (Bando et al., 2008). For a given neighbourhood around a pixel

, with eigenvalues of the covariance matrix trix

and covariance matrix diagonal elements ,, ,

This value is essentially how collinear the colour points in this neighbourhood are in RGB-space. These are visualized in Figure 3. The lower this value is, the more collinear the points are and the higher the less collinear. As mentioned in the appendix of the paper, this can be considered to be related to cross-correlation, and is thus exactly what we want to use. How to choose this neighbourhood size is a different story. The smaller it is, the less statistics we have about that neighbourhood and thus may have a worse matching. The larger it is, the better chance we have of a matching, however the longer it takes. This value is bounded between 0 and 1, as mentioned in the paper. If we show this for every pixel as in Figure 2, we can see that images without chromatic aberration show very little misalignment over the whole image. Another justification of using this metric is that according to the colour lines model (Omer and Werman, 2004), colour points in RGB-space of the whole image will lie along different colour lines. If we look at smaller neighbourhoods, then we can assume the points will also lie along either a line, or intersection of lines. This measures the collinearity of those neighbourhoods. If we search for an ideal alignment, then we want to maximize the collinearity, thus minimizing .

3.2. Finding Keypoints

The first task is to find regions where the keypoints would be useful. We want to find regions where the alignment measure is very high, but at the same time we want to be certain there is a good possibility for a good alignment. Choosing regions in the image with high values is costly, as we have to compute for every pixel and examine that image. Also, it is not guaranteed to give us a good pixel, such as smooth regions which may have ambiguous results. A good choice would be to find high gradient regions and use points from those. Specifically where we know there is an edge nearby. This gives us a better match since we know the other two channels should have a high gradient in that region too. We randomly sample from the norm of the gradient image with gradient sufficiently high within a threshold. In addition, if we want to pay the cost of the image, we can compute it only in regions where the norm of the gradient is sufficiently high. If we multiply by the norm of the gradient image, and threshold it, we can sample points that are highly unaligned with enough information around them to be aligned.

3.3. Finding Disparities

If we want to align two channels to the remaining one (say align green and blue to the red channel), we need to shift over all combinations of possible windows, with different scales. The idea is we want the neighbourhood of the green and blue channels to be correlated with the red, and those channels we know may be elsewhere with a different scale. Thus, we want to minimize the misalignment subject to a shifted and scaled window in both green and blue channels so we can write

With disparities , and relative scales . Where we iterate over all acceptable disparities (as in stereo) and all acceptable scales. We know however, that the disparities and cale difference shouldn’t be too large (unless one has a truely horrible lens) so we can limit the search to local neighbourhoods in that sixtuple. In fact, if the scale difference is decently small (which it usually is in the case of aberrations) we can simply look for disparities to find the scale aspect of the transform. Thus we can write as

To be able to handle different scales, one could simply do an image pyramid based approach as in other computer vision papers. This is unnecessary, as most aberrations are not that distant in the scale domain. The reason why this works is if we have a perfect edge (all one colour on one side, then all the other colour on the other side of the edge) in an image then the colours in the neighbourhood will cluster into two distinct regions. Since we are dealing with a natural image and edges are not perfect, these clusters will connect in a line as there will be a gradient from one colour to the other. If the image is misaligned, then this region will become more spread out. Now if we consider a multicoloured region, we get a more complicated shape. However, if we find disparities that minimizes this cluster’s collinearity, we should get a better aligned image since we’re minimizing the spread of the whole shape.

3.4. Pruning Keypoints

Although we’ve found keypoints and disparities for those keypoints in the other two channels, they might not give us good information. For example, if the best matching value for the point neighbourhood was high, we shouldn’t want to use that point as it is not a very well aligned point. We only want to use points that have gone from high value (unaligned) to low value (aligned). Since we only chose points that are unaligned, we just need to prune away the points that remain unaligned. Thus we only choose points with a low enough new

value. Alternatively, since we know we want to do a scale and translation for each channel (3 degrees of freedom for each) we only need 2 keypoints (each being 2 dimensional). Thus we could choose the 2 keypoints with the lowest

value. Other methods could include weighting the keypoints based on the parameter and haven’t been fully explored. In practice, thresholding by the right amount is sufficient for good results.

3.5. Computing Image Transforms

Now that we have a set of good points, we can solve for the transformation pretty easily. Let us consider the red and green channels for now. We have chosen in the fixed red channel and a point translated by disparity in the green channel . We have the equation:

Thus we can rearrange it for this point pair as:

If we have a second point we can solve this system for ,, , the scale, and translation respectively. In practice, we’d want more than just two points because the points might have only a good local solution and not global. Thus we’d want points from different regions of the image.

4. Results

(a)
(b)
(c)
Figure 5. LABEL: A synthetically aberrated mountain image via translation in different colour channels. LABEL: A corrected version using our algorithm. LABEL: The ground truth image.
(a)
(b)
(c)
Figure 6. LABEL: An image with exaggerated translation and scale shift in two colour channels. LABEL: A corrected version using our algorithm. LABEL: The same corrected version showing disparities and keypoints.
(a)
(b)
(c)
(d)
(e)
(f)
Figure 7. A set of extreme results. LABEL: An image taken with a lens with large chromatic aberrations. LABEL: A corrected version after using our algorithm. Similarly for LABEL: and LABEL:. LABEL: An image with distortion (Art and the Zen of Design, 2009) and our algorithm result in LABEL:. Note the properly aligned stand but misaligned horse head.

So far this seems to work quite well as seen in Figure 1 and more closely in Figure 4. In Figure 5 we have a synthetic image being corrected. As we can see the correction is quite similar to the original image. A lot of the high frequency details are a little blurred because the misaligned image has a lower resolution than the original image and thus we cannot get those details back. In Figure 6 we have a larger scale change between the different channels. Our algorithm solves this pretty well too. We can see the keypoints in the third image of this figure and some matchings aren’t always correct such as the one in the bottom left pointing in the wrong direction. In Figure 7, there are images with extreme lens distortion where our algorithm is expected to fail. Some of these include photos taken with a lens assembly that exaggerates chromatic aberrations. These photos are blurry because it is very hard to focus with this assembly. On the top we have the original photos and the bottom we have the corrected versions. Notice that the box is actually worse than the original whereas the tripod has a slight improvement. These photos are hard to deal with because the aberration is not as apparent since all of the colour channels are blurry.

5. Discussion

If we reduce the neighbourhood size to compute , we get more false-positives in the correlations. This is true because more disparities give a low number. The colours will cluster into a spherical region in a smooth neighbourhood, whereas we want lines. It will just choose a random disparity in this case. For the blurry images in Figure 7, this is a similar phenomena and there isn’t enough information in the image for this method to work well and reliably. Many disparities in this image regardless of direction gave a value below 0.01 where less than that is considered a ’good’ alignment in regions with more detail.

It was mentioned earlier that we may weight the keypoints and their appropriate disparities based on how much they reduced the alignment measure. Each row of equation 2 would be multiplied by a function of its associated . Different linear and squared error weighting based on have been attempted with little change in results. One could try to normalize the weights somehow instead of just using directly.

Also, there might be other statistical methods to explore to prune the keypoints such as computing the translation/scaling and getting rid of the outliers using RANSAC. The computationally expensive part of this algorithm is determining the disparities where we have a 4D loop (without handling extremely large scale changes). After we have the disparities it’s much quicker to deal with the keypoint data, especially since we need so few keypoints.

Another option to consider speedups is perhaps a hierarchical approach. One could try to solve the problem with a reduced resolution image and gradually work our way to the full resolution image using the lower resolution information.

What isn’t clear but is worth exploring is a statement made earlier saying that unwarping a distorted image will yield an image with a chromatic aberration that can be corrected with a scale/translation. This is worth exploring, as the undistortions relatively fast as well as this algorithm.

Another approach to find keypoints would be to segment the image into many cells and pick an appropriate keypoint from each as in section 3.3. This will allow enough information from different parts of the image to make a global warp more accurate than just choosing random points in the acceptable regions.

One other thing that has been explored is doing the same thing in gradient domain. Some initial results have shown that it didn’t work as well as the natural image formations as in our figures. It is worth exploring trying to align the gradients in a different way, perhaps using chamfer alignment on the edges of the channels.

6. Conclusion

We have presented an method that corrects chromatic aberrations in a single image without any use of calibration. Also, since this method is keypoint based and linear, the method is efficient. Our method also handles the case of saturation, since it can ignore those regions by not choosing keypoints near them.

Acknowledgements.
Wolfgang Heidrich for the idea for the project. Gordon Wetzstein for his lens.

References

  • Art and the Zen of Design (2009) External Links: Link Cited by: Figure 7.
  • Y. Bando, B. Chen, and T. Nishita (2008) Extracting depth and matte using a color-filtered aperture. In ACM SIGGRAPH Asia 2008 papers, pp. 1–9. Cited by: §2, §3.1.
  • T. E. Boult and G. Wolberg (1992) Correcting chromatic aberrations using image warping.. In CVPR, pp. 684–687. Cited by: §2.
  • W. Foundation (2009) External Links: Link Cited by: Figure 1.
  • J. Garcia, J. M. Sanchez, X. Orriols, and X. Binefa (2000) Chromatic aberration and depth extraction. In

    Proceedings 15th International Conference on Pattern Recognition. ICPR-2000

    ,
    Vol. 1, pp. 762–765. Cited by: §2.
  • S. B. Kang (2007) Automatic removal of chromatic aberration from a single image. In 2007 IEEE Conference on Computer Vision and Pattern Recognition, pp. 1–8. Cited by: §2.
  • J. Mallon and P. F. Whelan (2007) Calibration and removal of lateral chromatic aberration in images. Pattern recognition letters 28 (1), pp. 125–135. Cited by: §2.
  • I. Omer and M. Werman (2004) Color lines: image specific color representation. In Proceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2004. CVPR 2004., Vol. 2, pp. II–II. Cited by: §3.1.
  • R. Swaninathan, M. D. Grossberg, and S. K. Nayar (2003) A perspective on distortions. In 2003 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2003. Proceedings., Vol. 2, pp. II–594. Cited by: §2.
  • R. G. Willson and S. A. Shafer (1991) Active lens control for high precision computer imaging. Cited by: §2.