TE
科技回声
首页24小时热榜最新最佳问答展示工作
GitHubTwitter
首页

科技回声

基于 Next.js 构建的科技新闻平台,提供全球科技新闻和讨论内容。

GitHubTwitter

首页

首页最新最佳问答展示工作

资源链接

HackerNews API原版 HackerNewsNext.js

© 2025 科技回声. 版权所有。

How L1-minimization can recover highly degraded images

83 点作者 TrevorBurnham超过 15 年前

11 条评论

dmlorenzetti超过 15 年前
Only in the world of cookie-cutter journalism could this be described as a "chance discovery"-- a professor intentionally tests an algorithm, hoping to solve exactly the problem described, and discovers it works better than expected.<p>From the article: "Compressed sensing was discovered by chance... Candes... was experimenting with a badly corrupted version of the phantom meant to simulate the noisy, fuzzy images you get when an MRI isn’t given enough time to complete a scan. Candes thought a mathematical technique called l1 minimization might help clean up the streaks a bit... Candes expected the phantom on his screen to get slightly cleaner. But then suddenly he saw it sharply defined and perfect in every detail..."
评论 #1144722 未加载
gjm11超过 15 年前
For a much better explanation of this stuff from Terry Tao (mentioned in the article as being responsible for much of the theoretical basis of the technique; Fields medallist; very very smart chap) see <a href="http://terrytao.wordpress.com/2007/04/13/compressed-sensing-and-single-pixel-cameras/" rel="nofollow">http://terrytao.wordpress.com/2007/04/13/compressed-sensing-...</a> .<p>One important way in which I think the Wired piece is Just Plain Wrong: you <i>don't</i> sample a small fraction of the individual image pixels. You take a small number of samples, each of which is a kinda-random combination of image pixels.<p>Some examples and further quite-friendly explanation here: <a href="http://www.acm.caltech.edu/l1magic/examples.html" rel="nofollow">http://www.acm.caltech.edu/l1magic/examples.html</a> .<p>More from Tao: <a href="http://terrytao.wordpress.com/2008/12/05/the-uniform-uncertainty-principle-and-compressed-sensing/" rel="nofollow">http://terrytao.wordpress.com/2008/12/05/the-uniform-uncerta...</a> and <a href="http://terrytao.wordpress.com/2009/05/25/reflections-on-compressed-sensing/" rel="nofollow">http://terrytao.wordpress.com/2009/05/25/reflections-on-comp...</a> both of which are mostly links to more detailed things, well worth reading if you don't mind mathematics.
评论 #1145176 未加载
yummyfajitas超过 15 年前
Here is a more detailed explanation of how this works.<p>Imagine you have a pixel image, with missing data, and with noise. Imagine you know <i>a-priori</i> that the image was generated simply by placing a few white circles against a black background. Then data was lost/discarded, and white noise was added to the remaining data.<p>(The white circles are cross sections of the patient's blood vessels or some such feature (bile ducts, in this case), and contain a contrast agent. The contrast agent appears white on the MRI, everything else appears black. )<p>Now that you know the pixel image has such a simple form, you don't need to do fancy noise removal to clean it up. All you need to do is figure out where the white circles are. Once you know where the are, you can just redraw the image based on your calculated locations of circles instead of reprocessing the old one.<p>You can play the same trick in k-space, which gives you a potent MRI reconstruction algorithm.<p>In practice, this doesn't work as well for more complicated images such as brains or abdomens. There is just much more to draw, and the images cease to be sparse.
评论 #1144749 未加载
harpastum超过 15 年前
Seems really interesting, but I don't understand why you'd get rid of 90% of the image -- wouldn't it be much more useful to simply enhance the current images to 10x higher resolution?<p>That seems to be what the scientists are doing for the MRI, but the author seems to think this would be useful in that it would allow people to throw away a lot of data. I think it would be much better for extrapolating even farther from the data we have.
评论 #1144524 未加载
评论 #1144585 未加载
评论 #1144515 未加载
yaroslavvb超过 15 年前
There's some confusion between "compressed sensing" and "L1-minimization" in the comments and the article.<p>L1 minimization here refers to trying to represent the image as a linear combination of basis functions a1 f1+a2 f2+ ... by minimizing objective function fit(a1,a2,...)+sum_i |a_i| where "fit" represents how closely the reconstruction fits the image (for instance, squared distance). The second term in the sum is the L1 norm of the parameter vector, and it has a special property of being non-differentiable at 0, and often causing individual components a_i to become exactly 0 during minimization, hence resulting in "sparse" solution vector.<p>Compressed sensing refers to the idea that when the image is sparse in the chosen basis {fi}, it's sufficient to take a small number of random measurements of the image. We can then compute "fit" approximately which gives us a different objective function that we minimize. The result will be close to what we'd get if we used the original (full) "fit" function
fadmmatt超过 15 年前
Enhance!<p><a href="http://www.youtube.com/watch?v=Vxq9yj2pVWk" rel="nofollow">http://www.youtube.com/watch?v=Vxq9yj2pVWk</a>
评论 #1145180 未加载
gane5h超过 15 年前
Cool. Didn't expect so much interest on HN for a highly mathematical concept.<p>If you're interested in a blog solely focussed on CS, here's one: <a href="http://nuit-blanche.blogspot.com/" rel="nofollow">http://nuit-blanche.blogspot.com/</a>
评论 #1145818 未加载
评论 #1144994 未加载
iskander超过 15 年前
I might be mistaken (I'm not an expert), but isn't the article mixing up compressed sensing and low-rank matrix completion?<p>edit: Also, they present all of compressed sensing as if it were just that one algorithm for image reconstruction (which samples in the original space, rather than an incoherent basis). Tech media fail?
评论 #1147258 未加载
tocomment超过 15 年前
Why hasn't this made it into Photoshop or GIMP yet? What would it take to implement this to enhance images?
评论 #1144720 未加载
评论 #1146549 未加载
hammmatt超过 15 年前
I'm curious about what are the consequences of the discovery.<p>What effects will this have on the storage of photos?<p>Can this principal be extrapolated to other forms of data?
评论 #1144985 未加载
RiderOfGiraffes超过 15 年前
More details: <a href="http://news.ycombinator.com/item?id=756176" rel="nofollow">http://news.ycombinator.com/item?id=756176</a>