Edgar (ECE): Hello, welcome to my presentation, Optimal Coding in Spatial Spectral Compressive Spectral Imaging. I know this title sounds a little bit fancy, but we are going through the details in order for you to understand this topic. So let's start Edgar (ECE): with defining Spectral Imaging. Spectral imaging is basically capturing the spatial and the spectral information of a given scene. Edgar (ECE): The spectral information is basically the electromagnetic radiation or spectrum, that is, the colors; there are many ways in order to do this Edgar (ECE): imaging process, the first one is, for example, the point scanning, that you can see, on the left part of my presentation; in the point scanning Edgar (ECE): you get the spectral information point by point of a given scene. This will be, of course, time consuming. The other one is the line scan where you pass line scanner through the scene, and the other one is basically Edgar (ECE): wavelength scan in which you filter the information, wavelength by wavelength; so the main drawbacks of these Edgar (ECE): approaches is, for example, the time; it will be really time consuming, and other different drawbacks. Edgar (ECE): The question is how to capture all this information, using a single snapshot, single snapshot, so we are going to to try to answer this, over this presentation. Edgar (ECE): But first let's talk about some applications; the spectral imaging is used for example in food quality, in art conservation in medical imaging, Edgar (ECE): and even in remote sensing, and in all of these scenarios the spectral imaging is used in order to extract hidden information to our eyes, for example in the art conservation you can see some hidden patterns in this drawing, so this is the key concept, hidden information. Edgar (ECE): Let's talk about how to capture this information, using a single snapshot, Edgar (ECE): and moreover using a grayscale sensor Edgar (ECE): So, if you want to do that, something has to happen before, right, and that's something is the CODED APERTURE. Edgar (ECE): So for the ones who knows about communication imagine two users are trying to use the same channel, at the same time; if the receiver wants to know Edgar (ECE): which user is actually on the channel Edgar (ECE): the user has to have some code, right, and in this case, something similar is happens, but we are not talking about users, but we are talking about spectral bands. Edgar (ECE): We're not talking about receivers but we're talking about grayscale sensors, but the concept is quite of the same; this slide shows the Coded Aperture Snapshot Spectral Imager (CASSI), which is a state of the art architecture in Compressive Spectral Imaging. Edgar (ECE): There are many variations of the CASSI and we're going to talk about the SSCSI or SPATIAL SPECTRAL COMPRESSIVE SPECTRAL IMAGER. Edgar (ECE): Two main variations here with respect to the CASSI, the first one is that the coded aperture is going to be located after the spectral dispersion, and the second one is that Edgar (ECE): the scene is going to in focus, or on focus at the sensor. So as you see, the coding process which is key here is going to depend on the coded aperture position with respect to the sensor. Edgar (ECE): So let's do a little comparison between CASSI, and SSCSI, TOP VIEW. In the CASSI, the coding happens before the spectral dispersion and a spectrally dispersed image is going to arrive to the sensor. In the SSCSI, on the other side Edgar (ECE): the coding is happening after the spectral dispersion and an in focused image is going to impinge on the sensor. Edgar (ECE): So we did two contributions here; the first one was trying to fully characterize the SSCSI; that is coming from analog to digital, creating a digital model. Edgar (ECE): So, in order to do that, we have to define both experiments Edgar (ECE): and in theory the spectral resolution and the spatial resolution in X and Y. The spectral resolution Edgar (ECE): is going to depend on three Edgar (ECE): elements, so the first one is the coded aperture position with respect to the sensor; the second one is the coded aperture pitch size, and the third one is the dispersion process. Edgar (ECE): On the other side, the spatial resolution in X is going to depend on the coded aperture position, Edgar (ECE): the coded aperture pitch size, and the detector pitch size, and the third one, which is the spatial resolution in the Y domain, is going to depend on the coded aperture and the detector pitch sizes. This was our first contribution. Edgar (ECE): Our second contribution was related to the optimality on the patterns; remember the patterns, or the mask, or the coded aperture are key Edgar (ECE): here in compressive spectral imaging. If you want to go from sensor measurements to a hyperspectral image, you have to apply some algorithms Edgar (ECE): And the performance of the algorithm is going to depend on several things, for example, the sparsity of the scene, that is, how many zeros are on that scene Edgar (ECE): the number of snapshots that you capture, you can capture more than one snapshot; the calibration process and, of course, the structure of the coding patterns; we are focused on the structure of the coding patterns Edgar (ECE): So we found. Edgar (ECE): some interesting results here in our research. The first one is that the optimal coding patterns in the SSCSI are not going to be fully random. Edgar (ECE): The random patterns are usually implemented, but the optimal are not fully random. Edgar (ECE): And the second one is that the structure of the coded aperture is going to depend on the coded aperture position. Here on the left, you see a random pattern and on the right you see an optimal pattern when the coded aperture is on top of the sensor. Edgar (ECE): And here, you see the optimal pattern when the coded aperture is a little bit away from the sensor; you see that both coded apertures, they have structure but different structure. Edgar (ECE): You see here the results, this is the original hyperspectral scene, Edgar (ECE): and this one is the optimized . You see that the closer reconstruction is going to come from the optimized coded apertures, if you compare it to random, and this boolean is another kind of structure Edgar (ECE): To summarize the optimality in the patterns, we found two different conditions, so the first condition is that the ON pixels or the ONE pixels or the white pixels are going to be Edgar (ECE): are separated as possible from each other, so this is one condition, and the second condition is that they are going to be some clusters on the rows, depending Edgar (ECE): On the coded aperture position. Here, you see, for example, the clusters but here you don't see any cluster. Edgar (ECE): So final thoughts as a final Edgar (ECE): conclusion we, we can say that the quality of the images, the recorvered hyperspace scene is going to depend on several factors. I am going to mention some of them. It is going to depend on the calibration, it is going to depend on the quality of the lenses, on the dispersion process, and also the coded aperture pattern. Edgar (ECE): and Edgar (ECE): Our second main conclusion is that the coding mask is a key factor in any compressive spectral imager. Edgar (ECE): So this is the reference we use for our presentation, and this is the result Edgar (ECE): in bibliography that we had in our research, so we published some conferences and some Edgar (ECE): full journals Edgar (ECE): I think that would be all for this presentation, thank you very much for your time and I will see your questions session, thank you.
Optimal Coding in Compressive Spectral Imagers, Edgar Salazar
From Huma Rasheed April 14, 2021
18 plays
18
0 comments
0
You unliked the media.
The spatial spectral compressive spectral imager (SSCSI) is a proposed optical architecture, where the spatial-spectral coding of a scene is realized by a binary ON-OFF mask and a dispersive element. The quality of the recovered hyperspectral scene in SSCSI depends on various parameters, such as the coded aperture location with respect to the sensor, the coded aperture and detector pitch sizes, and the diffraction introduced by the grating. Given these parameters, the sensing matrix of the imager is largely dependent on the coded aperture structure and characteristics. As such, its optimization is critical to maximizing the quality of the recovered scene.
- Tags
- Department Name
- Electrical and Computer Department
- Appears In
Link to Media Page
Loading
Add a comment