Sunday, October 23, 2011

You keep using that word

I got interested in reading this paper [1] because of the title, alas, the authors keep using that word but I do not think it means what they think it means.Here is the interesting bit [italic is my emphasis]

Recent work has investigated the concept of compressive sensing for imaging applications. Compressive sensing is similar in concept to image compression except that it reduces the number of samples acquired rather than remove redundancy from the image post-acquisition. In an example of this Robucci et al describe an imaging system that uses a custom sensor to perform ‘noiselet’ transforms (a form of pseudo-random wavelet transform) during image acquisition [14]. This is an interesting approach, but it requires specialist sensor hardware that is not yet commercially available and primarily concentrates on image, rather than video, acquisition.

In this paper, we present a camera system that attempts to overcome the limitations of these strategies by utilizing a form of compressed imaging. The camera system adaptively concentrates most of the available bandwidth on a subset of pixels while sampling the remainder of the pixels at low-speed. In this way the camera is able to use most of its available bandwidth to capture images of the ROI at a relatively high-speed while simultaneously acquiring high-resolution images at low-speed using the rest of the bandwidth. In addition, the camera is able to use this full-frame information to calculate the positions of targets and update the high-speed ROIs without interrupting acquisition. This allows the camera to track and image moving targets at high-speed while simultaneously imaging the whole frame at a much lower frame rate. While the overall bandwidth of the camera is low it is utilised efficiently and this allows changing ROIs to be acquired continuously at high speed. Because the total volume of data is reduced considerably by only collecting the necessary information at the source, pressure is reduced on down-stream systems; including real-time data analysis.

Let me first emphasize something, this implementation of a camera is great but there is absolutely no need to capture the wording compressive sampling in it when the work stands on its own. The initial and most important aspect of compressive sensing is that the whole field of the image is sensed all the time. Let us be more specific:

Tracking high-speed ROI with low-speed full-frame.
Finally, after simultaneously acquiring the ROI at high speed and the full-frame at a lower speed, the full-frame image can be analyzed to determine the most appropriate ROI (Figure 9). The Analysis loop can then pass any changes to the ROI addresses to the Acquisition loop. The Acquisition loop gets the most recent set of ROI addresses from the FIFO buffer when it is ready and so continuous acquisition is not interrupted. In this way moving targets can be tracked.

Here the ROI updating goes as slow as the low-speed full frame acquisition. This is where the process described in this paper cannot be described as compressive sensing. In CS, the ROI (in the case of adaptive compressive sensing) is updated at the full speed of the acquisition. So let me restate, this is a great implementation but this is not a compressive sensing implementation.

[1] A Reconfigurable Real-Time Compressive-Sampling Camera for Biological Applications, Bo Fu, Mark C. Pitter, Noah A. Russell.


Fool said...

Hi there,
Thanks for your sharing and comments. This is Bo, the author of this work. I saw your comments on PLoS ONE and found your blog via your twitter (sorry for that:p).

'Compressive sampling' here happens on the data acquisition level. So we analyzed the whole frame and we considered the procedure of finding the ROIs as 'compress'. Then sample these regions with most of the B/W.

The picture in ROIs are being updated with most of the B/W, but the 'locations' of them are not updated that fast. So the 'tracking speed' is not a selling point of this camera.

I agree with your comments. This is not a usual way of using the term 'Compressive sampling' and I'm sorry if it was misleading. I can discuss about this with my supervisor:)


Igor said...

Hello Bo,

Let me reiterate that I think this is a great implementation you have here.

Also, please bear in mind that nobody owns a trademark on the wording "compressive sensing".

However, what you are doing is really called adaptive sampling. It assumes that no large changes will occur in between slow frames.

Unlike a compressive sensing (adaptive or non adaptive) imager, your system does not sense the full FOV at high frame rate. It only sense the full FOV at the lower frame rate and only locally at the high frame rate. A CS imaging system would sense the full FOV at a high frame rate.

Your imager could become a CS system if the ROI pixels were providing information about the full frame. For that, you would probably need some other sort of device in between the field of view and the CMOS but that would be a whole different paper.


Bo said...

Dear Igor,
Thank you for your reply. And I agree with your words. Adaptive sampling seems more suitable for this work.
I'll add that to my thesis:) Cheers.