K-L Reconstruction of the heteroclinic mode C test on 64x64 images



click here (xxx) to view an mpeg movie of the flame motion.



To perform the K-L analysis, videotape of the flame front was digitized at a rate of 30 frames/second. 100 images of dimension 64x64were used. The average of the image ensemble (denoted by I0) was subtracted from each image prior to the K-L decomposition. The K-L analysis produced a set of eigenvectors that are denoted by I1, I2, ... where I1 represents the K-L eigenvector containing the most energy. The resulting K-L energy spectrum is shown at the left. The spectrum indicates that a K-L basis containing the first 2 vectors contains 75% of the energy.



The ensemble average I0 and the 20 K-L eigenvectors are:

The images in the original data set are denoted by U(t) for t = 1 .. 100. The m-th order K-L reconstruction of image U(t) is:

Um(t) = I0 + a1(t)I1 + a2(t) I2 + ... + am(t)

where ai(t) is the projection of U(t) on Ii. The successive reconstructions for U(1) are shown below (left to right top to bottom). The first image is the average of the ensemble and the last image is the original image from the dataset.



An animation showing a side-by-side comparison of selected K-L reconstructions can be found at:

The constructions from left to right are order-1 K-L reconstruction, order- 2 K-L reconstruction, order- 4 K-L reconstruction, and the original image.

This view of the heteroclinic mode C test on 64x64 images state consists of 3 dimensional renderings of selected K-L reconstructions. The upper left corner is the order-1 K-L reconstruction, the upper right corner is the order- 2 K-L reconstruction, the lower left corner is the order-4 K-L reconstruction and the lower right corner is the original image. The animations are:



In order to emphasize the relative contributions of the different K-L modes to the motion as a function of time, this animation displays an instantaneous bar chart of the projections of the images on the first 20 K-L eigenvectors. The first bar represents the instantaneous size of a1(t), the second bar represents the instantaneous size of a2(t) and so on. The middle image is a 3-dimensional rendering of the K-L reconstruction using the average and the first 20 K-L eigenvectors The rightmost image is a 3-dimensional rendering of the original data. The animations are:



This animation shows a bar chart of the instantaneous values of the projections of the images on the first 20 K-L eigenvectors. The resulting reconstruction is presented as a grey scale image (middle figure) and is compared with the original image (rightmost figure). The animations are:



The following animation more clearly separates the contributions of the modes. Each row corresponds to a reconstruction. The leftmost image in each row is a bar chart showing the instantaneous amplitudes. The right image shows the corresponding reconstructions. The animations are:



These animation show the correlation between the 3D phase plot of selected projections, the appearance of reconstruction and the original image. The phase space plot in the animation shows 1 vs 2 vs 3 The middle image shows the order- 3 KL reconstruction, while the rightmost image shows the original image. The animations are:



This animation shows different views of phase plane plots of the projections. In each phase plane plot, all of the projections are represented as black dots. The current position in phase space is marked by a larger red dot. The animations are laid out:

a 1(t) vs a 2(t) a 3(t) vs a 4(t) a 1(t) vs a 3(t)
a 1(t) vs a 5(t) a 3(t) vs a 5(t) a 4(t) vs a 5(t)
a 1(t) vs a 7(t) a 3(t) vs a 7(t) a 7(t) vs a 8(t)