Over the past two months, I’ve been running a number of classifiers against Mena’s method, using various fusion techniques. This resulted in so much work that my supervisor and me decided to create two journal articles, rather than one.
I only need to run one more experiment, after which I will give a sneak peek of my findings on this site.
(I’ll be away on leave next week though)
It’s been a while since my previous post, but I have been making a lot of progress. Currently busy with a paper for
ISPRS – International Society for Photogrammetry and Remote Sensing.
I’ll post my findings and progress up to date once I’ve submitted the paper.
As I mentioned in an earlier post, I investigated a number of different colour spaces to see if that might imporve my results. Just some quick feedback on that:
I have not used a quality metric to quantify my results, just a quick visual comparison.
The fact that RGB seems to be the best colour space for BHTCC contradicts Mena and Malpica’s findings that HSI is better than RGB. This leads me to believe that there is still something buggy with my code…. bleh. Must have double checked that code a million times now. Perhaps I’ll get someone else to look at it for me.
The final process in Mena and Malpica’s low-level classification is to fuse the outputs from the 3 classifiers (Mahalanobis, Bhattacharyya and Bhattacharyya Haralick Texture Cube). Jeez, these names are crazy. To top it all off, my surname is Hauptfleisch. So once i’ve made some changes it will be the Bhattacharyya Haralick Texture Cube Hauptfleisch method. Hehehe… crazy.
Anyway, back to the fusion process. Dempster-Shafer’s Theory of Evidence [1,2] was used to fuse the results from the 3 classifiers. Thus, I’m implementing an OSSIM component that will fuse 3 grayscale images.
[1] A. P. Dempster, “A generalization of Bayesian inference,” Journal of the Royal Statistical Society, vol. Series B, no. 30, pp. 205–247, 1968.
[2] G. Shafer, A Mathematical Theory of Evidence. Princeton University Press, 1976.
Not being entirely convinced that RGB is the best colour space to do mathematical computations in, I’ve started running some of my algorithms in the HSI colour space. Even though I did not do exhaustive testing, the HSI space did not yield very promissing results.
One reason might be because HSI and HSB have a discontinuity between 0º and 360º. When computing distances between values, which is the basis of a number of my classification algorithm, one might get inaccurate results.
Figure 1: HSB Colour Space
In Figure 1, it is clear that the red colours at 0º and 360º are quite ‘close’, i.e. a distance measurement should be close 0. However, because of the discontinuity a big distance is obtained between the yellowish-reds and the pinkish-reds.
Another colour space one might consider is the CIE 1931 space, which was the first mathematically defined colour space. However, it is not suitable for giving direct difference between two colours.
The Lab colour spaces are later renditions of mathematical colour spaces, which allows one to more accurately measure the distance between two colours. These measures are (apparently) not perfect, but it seems to produce the best results at present.
There’s a pretty good article on colour spaces on CompuPhase’s website, which provides insights into the difficulties involved with colour metrics. It is also suggested to use the square root of L* in Lab, rather than standardized cube root.
Next thing would be to implement the Lab conversion and to do a couple of experiments.
Its a brand new year! All the best to all in your ventures.
I’ve made some progress since 2006. It looks like the Texture Bhattacharyya-Distance Classifier (what a mouth full) is finished. I tested it on some synthetic images and the results are as expected. Here’s some results:
Training Sets:
 |
 |
The textures in both Figures 1.1 and 1.2 is exactly the same. The only difference is that the texutre in Figure 1.1 is accross the RGB bands, whilst 1.2 is spatially accross the pixels.
Test Sets:
 |
 |
The top right blocks in these test sets, are the same as the training sets. Thus, the resulting image should have the lowest (black) values in the top right corners.
The top left and bottom right blocks should be the same distance from the training set, but in different directions.
The bottom left block is the most dissimilar from the training sets, so one would expect it to be the brightest (white) of all 4 blocks.
Results:
 |
 |
The results are as not quite as expected with:
Thus, either there is something wrong with my implementation, this classification method simply doesn’t work or I am using the incorrect colour space.
Real-World:After running the classifier on somereal-world data, the results are nowhere near as pleasing as expected. Perhaps, because I did not fully understand the workings of the algorithm. I’m doing closer analysis of the real-world results, trying to understand the algorithm a bit better. Will post the results once I’ve managed to figure out how to improve or to explain the results.
I reckon this will be my final post for 2006.
Things have really been quite busy the last couple of weeks. I’ve managed to do quite a bit coding and I my 3 classifiers from Mena and Malpica are nearing completion. I only need to debug the final classifier a bit… and try to speed it up a bit. It’s absolutely crawling!!
As usual, I’ll put up some results once finished.
I’ve also uploaded our poster paper presented at the PRASA 2006 conference. In the paper, we implemented the ACE method by Doucette et al. and replaced the Canny edge detector with a Neural Network based edge detector. The results? There was no significant difference between the two approaches, even though the Neural Network based classifier did yield more visual pleasing edges.
I finally managed to upload the results from Mahalanobis and Bhattacharyya distance pixel classifiers. These results are the first obtained and have not been optimized yet. I’m sure I’ll be able to improve the results. Once I finish the third and final classifier, I’ll start optimizing the parameters and training sets.
Here’s the results in the following order for three different scenes: original images, Mahalanobis distance classifier and Bhattacharyya distance classifier.
Things have been pretty crazy the last couple of weeks. I’ve been rushing to get a paper finished for the Pattern Recognition Association of South Africa (PRASA) conference. Submission deadline is Wednesday, so hopefully I’ll be able to blog on some of the results thereafter. I’ll post the PRASA work, as well as Mena and Malpica’s stuff.
I’ve managed to complete the first two pixel classifiers suggested by Mena and Malpica (2005) and the results are quite good.
Since the outputs from the classifiers are merely distance values for each pixel, it is not really suitable for visualization. I will, however, try to create some enhanced images and post them here. Everyone likes pictures!!
