Hyperspectral imageexample

Read this 2023 press release about imec's combined VIS & NIR spectral camera system, complemented with hi-res RGB imaging, for data acquisition at video rate – allowing flexible assessment of the pros and cons of different spectral resolutions and ranges.

Our unique on-chip hyperspectral imaging technology allows us to make different spectral filter patterns. Those fall into two categories:

$$ V(x) = \begin{cases} \overline{V}(x) + c_1 & x = 0 \\ \overline{V}(x) + c_1 - p_0 & 0 < x < x_1 \\ \overline{V}(x) + c_1 - p_0 - p_1 & x_1 < x < x_2 \\ \overline{V}(x) + c_1 - p_0 - p_1 - p_2 & x_2 < x < L \\ \overline{V}(x) + c_1 - p_0 - p_1 - p_2 - p_L & x = L \end{cases} $$

Hyperspectral imagemeaning

Traditionally, different approaches are taken to solve a beam deflection problem depending on the particular anchors present. However, in order to simplify the implementation, we have considered all beams to be free-floating with the same set of boundary conditions. This establishes a base set of equations and unknowns with zero degrees of freedom. Then, each anchor adds either one or two additional equations and unknowns in a predictable way. This allows us to use one implementation to solve any configuration of anchors.

If we add a fixed joint to the left end of the beam, we add two additional unknowns and equations. The unknowns are $p0$ and $m0$, and the equations are $y(0) = 0$ and $\theta(0) = 0$:

At this point, we recognize that the constants $c_1$ and $c_2$ must both be equal to zero in order to satisfy our first two base equations, $V(0) = 0$ and $M(0) = 0$. We also make a few other simplifications:

The shear force at $x = x_1$ and $x = x_2$ is undefined due to the discontinuities at these points, but this doesn't matter since calculating the integral across a discontinuity is trivial.

Traditional hyperspectral imaging scanners contain a lot of precision optics to select and diffract the light. This makes them relatively heavy, expensive, slow and delicate – with a need for frequent recalibration.

The numerical integration to find $ybar$ is only half the battle. To fully solve the problem and find the as yet unknown constants of integration, consider this generalized example. Assume that we have a fixed anchor on the left with reaction force $p_0$ and moment $m_0$, a fixed anchor on the right with reaction force $p_L$ and moment $m_L$, and two pinned joints located at $x_1$ and $x_2$, with unknown reaction forces $p_1$ and $p_2$. The presence of anchors means that the integration is carried out piecewise. The shear force $V$ is then given by:

We've taken a bit of a leap here moving from five to three piecewise portions, which one can confirm by working out the various pieces, or simply come to accept by recognizing that $\theta$ must be continuous.

$$ M(x) = \begin{cases} \overline{M}(x) & x = 0 \\ \overline{M}(x) - p_0 x + m_0 & 0 < x < x_1 \\ \overline{M}(x) - p_0 x - p_1 (x - x_1) + m_0 & x_1 < x < x_2 \\ \overline{M}(x) - p_0 x - p_1 (x - x_1) - p_2 (x - x_2) + m_0 & x_2 < x < L \\ \overline{M}(x) - p_0 x - p_1 (x - x_1) - p_2 (x - x_2) + m_0 + m_L & x = L \end{cases} $$

Because the constants of integration and anchor forces and moments are unknown, we have to carry them symbolically through the integration. In practice, we begin by first numerically integrating only the known loads on the beam, and add the unknown forces after the integration. We will use an overbar to show that we have ignored the unknown constants of integration and anchor loads and moments:

Likewise, adding a fixed joint to the end of the beam does not alter the original $M(0) = 0$ boundary condition, since an applied moment causes a discontinuity in $M$, and the boundary still holds on one side of the discontinuity.

Imec can assist you in tailoring its hyperspectral imaging technology to your application. For instance, why not use one of our high-resolution cameras  to define which frequency bands are relevant for your needs? Afterwards, we can help you to build your dedicated filters or camera system.

$$ M(x) = \begin{cases} \overline{M}(x) + c_1 x + c_2 & x = 0 \\ \overline{M}(x) + c_1 x + c_2 - p_0 x + m_0 & 0 < x < x_1 \\ \overline{M}(x) + c_1 x + c_2 - p_0 x - p_1 (x - x_1) + m_0 & x_1 < x < x_2 \\ \overline{M}(x) + c_1 x + c_2 - p_0 x - p_1 (x - x_1) - p_2 (x - x_2) + m_0 & x_2 < x < L \\ \overline{M}(x) + c_1 x + c_2 - p_0 x - p_1 (x - x_1) - p_2 (x - x_2) - p_L (x - L) + m_0 + m_L & x = L \end{cases} $$

Hyperspectral imagesoftware

Each step introduces an unknown constant of integration. These unknowns are solved for based on the constraints applied to the beam, such as the presence of pinned or fixed anchors. Other unknowns, such as reaction forces from anchors, are solved for in the same way.

Beam Calculator solves for the deflection of a beam through direct integration. This is done in four subsequent integrations, which solve for the shear force ($V$), moment ($M$), slope ($\theta$), and deflection ($y$) of the beam:

Now say we add a pinned joint $p1$ at some location $x1$ on the beam. This introduces one additional unknown, $p1$, the force applied on the beam by the joint. It also introduces another equation: $y(x1) = 0$. There are still zero degrees of freedom:

Hyperspectral imageprocessing

Another way of looking at a hyperspectral image is as a set of image layers, each in another wavelength of the electromagnetic spectrum. The combination of those layers is called a hyperspectral data cube.

Hyperspectral imageclassification

$$ \theta(x) = \begin{cases} \overline{\theta}(x) + \frac{1}{EI}\left(-p_0 \frac{x^2}{2} + m_0 x\right) + c_3 & 0 \le x < x_1 \\ \overline{\theta}(x) + \frac{1}{EI}\left(-p_0 \frac{x^2}{2} - p_1 \frac{(x - x_1)^2}{2} + m_0 x\right) + c_3 & x_1 < x < x_2 \\ \overline{\theta}(x) + \frac{1}{EI}\left(-p_0 \frac{x^2}{2} - p_1 \frac{(x - x_1)^2}{2} - p_2 \frac{(x - x_2)^2}{2} + m_0 x\right) + c_3 & x_2 < x \le L \\ \end{cases} $$

At the present time, the implementation requires the beam to have constant rigidity $EI$. This is so that unknowns which appear in intermediate integrations can be carried through symbolically, and all solved simultaneously at the end.

To perform the numerical integration and calculation of $\overline{V}$, $\overline{M}$, $\overline{\theta}$, and $\overline{y}$, we use the trapezoid rule and Simpson's rule, discretizing the functions on a grid of points between 0 and $L$. We also add a grid point at each point load or anchor. During the integration, point loads can create discontinuities. In these cases, two grid points are created with the same x-coordinate, and each is assigned to one side of the discontinuity.

There are multiple hyperspectral imaging technologies. Each applies different techniques to filter the light and capture the image data. The key differentiators you need to keep an eye on are:

Of course, we can also develop custom hyperspectral image sensors, with your desired pattern lay-out, filter specification, imager chip, ... All the way up to the design of the complete camera system.

In the last two integrations, the number of grid points at which an exact integral can be calculated is halved as a result of using Simpson's rule. Using Simpson's rule is required because we are integrating a higher-order function. The trapezoid rule is sufficient for the first two integrations because it can still produce exact numerical integrals of zero- and first-order functions. This also means that the applied load on the beam must be evaluated at more grid points than will be present in the final result for ybar.

We have now worked out the expression for the fully generalized beam. Adding additional pinned joints just increases the number of piecewise portions.

Selecting the required properties is a tradeoff, driven by the requirements of the application. If you know which spectral bands to look for, there’s no need to capture the full spectral range. Especially because a lower resolution will enable faster acquisition speeds and open new possibilities with hyperspectral video imaging.

Any number of joints can be added in this way, and the degrees of freedom remain unchanged. In practice, only fixed joints on the ends of beams are allowed, because a fixed joint in the middle of a beam would divide it into two independent beams, each having a fixed joint on one end.

The base set of equations is obtained by considering an unsupported beam. At each end of the beam, the shear force and moment are both zero, since there are no applied moments or forces. For this trivial problem, we have 4 unknowns (the 4 constants of integration) and 4 equations: $V(0) = 0$, $M(0) = 0$, $V(L) = 0$, and $M(L) = 0$.

$$ y(x) = \begin{cases} \overline{y}(x) + \frac{1}{EI}\left(-p_0 \frac{x^3}{6} + m_0 \frac{x^2}{2} \right) + c_3 x + c_4 & 0 \le x < x_1 \\ \overline{y}(x) + \frac{1}{EI}\left(-p_0 \frac{x^3}{6} - p_1 \frac{(x - x_1)^3}{6} + m_0 \frac{x^2}{2}\right) + c_3 x + c_4 & x_1 < x < x_2 \\ \overline{y}(x) + \frac{1}{EI}\left(-p_0 \frac{x^3}{6} - p_1 \frac{(x - x_1)^3}{6} - p_2 \frac{(x - x_2)^3}{6} + m_0 \frac{x^2}{2}\right) + c_3 x + c_4 & x_2 < x \le L \\ \end{cases} $$

Adding a pinned joint to the end of the beam does not alter our original $V(0) = 0$ boundary conditions, so long as we account for the fact that a point load causes a discontinuity in $V$, and that on one side of the discontinuity, the boundary condition will still hold.

Hyperspectral imagedownload

When a point load is added to a Beam, it is frozen and cannot be mutated. If you need to change a point load, first remove it using removePointLoad, and then add a new point load.

Our filters eliminate the complex optical camera design and can be deposited on a commercial CMOS imager, like the one in your smartphone or a model with scientific imaging capabilities.

The final step is to solve for the remaining unknown variables $c_3$, $c_4$, and others introduced by the various anchors. The remaining boundary conditions not yet considered are $V(L) = 0$ and $M(L) = 0$. Each pin joint adds one unknown $p_i$ and one equation $y(i) = 0$, and each fixed joint adds the unknowns $p_i$ and $m_i$, and the equations $\theta(i) = 0$ and $y(i) = 0$. Thus, the degrees of freedom for the problem is always zero. This is a straightforward linear algebra problem that can be solved with a method such as LU-decomposition. The difficulty lies in preparing the matrix and keeping all our ducks in a row.

Hyperspectral imagedataset

For a beam of constant rigidity $EI$, the Euler-Bernoulli equation describes the relationship between the deflection of the beam and the applied load:

This website offers you an overview of our ready-to-use evaluation kits, including remote support, and the easy-to-use HSI STUDIO, HSI MOSAIC and HSI SNAPSCAN software suites to make hyperspectral imaging accessible to anyone.

Hyperspectralimaging camera

Whenever possible, we would like the numerical integration to be as accurate as possible. It turns out that if the applied load to the beam, $w(x)$, is either uniform or a point load, we can carry out each numerical integration exactly. Here's why:

Imec’s hyperspectral imaging technology takes a different approach. As a world-leading R&D hub in nanotechnology, we have developed a wafer-level CMOS process to integrate thin-film spectral filters directly on the pixels of the image sensor.

Hyperspectral imaging technology is designed to make images that reveal a maximum of spectral information for each pixel of the image.

We won't know a priori if the applied load is a constant or not--the user simply provides a function. We think the method of integration above will provide sufficient accuracy for non-uniform loads, and has the benefit of providing exact answers for uniform loads.