Parametrization and Machine Learning from Dark-Field Microscopic Images of Dried Drops of Water
Matjaž Zadnik, Igor Kononenko, Matjaž Bevk
University of Ljubljana
Faculty of Computer and Information Science
Tržaška 25, 1000 Ljubljana, Slovenia
ABSTRACT
We use the dark-filed microscopy to record images of dried drops of water. We use this approach in order to visualize the effect of a natural source of subtle energy, which can be found in village Tunjice near Kamnik in Slovenia, on various water solutions. We tried to detect differences in images of control and experimental water. The images were described with various sets of numerical parameters and various machine learning algorithms were run in order to detect dfferences. Besides, each numerical parameter was tested for statistical differences between the two groups of images. The results show that for the solution of Aluminium Oxide the differences are detectable while for ordinary tap water, filtered water with reverse osmosis and for spring water from Tunjice we were not able to detect any differences.
- Introduction
Is water capable of storing information? We tried to explore this phenomena by analyzing the patterns of different water solutions after being exposed to subtle natural energy field that can be found in village Tunjice near Kamnik, Slovenia. The method of observing the dried drops of liquids was discovered by Ruth Kuebler from Stuttgart [4]. She began examining and photographing dried water droplets under the
dark-field microscope. She discovered that the dried drops of water reflected typical structures according to their origin. The effects of magnetic fields, radioactive waves and different types of water were tested. The investigations were extended through the work of Minnie Hein [4], who concentrated on the pictures of dried drops of various types of essences and human saliva. Our preliminary experience with this method is described in [3].
The paper is organized as follows: In the next Section we give a brief description of the experimental method. Section 3 contains a brief description of used methods for parametrization of images. The fourth Section describes different machine learning algorithms and results obtained with them. Section 5 gives the results of statistical analysis of all the numerical parameters and the last Section concludes.
- Recording of images of dried drops
The procedure was as follows. First the sus-
pensions were prepared:
-ordinary tap water
-filtered tap water with inverse osmosis
-spring water from Tunjice
- 0.01% of Al2O3 in destilled H2O
Then each suspension was divided into an experimental and a control glass. The experimental glasses were then exposed for three days to the natural source of subtle energy. The next task was to make drop samples of the suspensions. Drop samples were made with a sterile syringe on a clean microscopic slide. The average diameter of a drop was about 5 mm. At this point the drops were dried before further processing. To take pictures of dried drops, we used a dark field microscope with a mounted digital camera. A relatively small magnification factor of 40 and 100 was used. Each drop separately was observed under microscope and photographed. Some images had to be eliminated due to inappropriate size of a drop or due to inappropriate shape.
The size of images was 2272 x 1704. The Color depth was 24 bit. In the next step we extracted an inner rectangular area of magnitude 500x500 for each drop.
- Parametrization of images
For description of images with a set of numerical parameters we used three different approaches.
3.1Statistical texture description
Our description consisted of first and second order statistics [13, 1]. First-order statistics are computed from a function that measures the probability of a certain pixel occurring in an image. This function is also known as histogram. We can regard this function as a probability function of pixel values and we can characterize its properties with a set of statistical parameters (also called first order statistics) [6]. Second-order statistics operate on probability function, that measures the probability of a pair of pixel values occurring some vector apart in the image. This probability function is also called cooccurrence matrix, since it measures the probability of cooccurrence of two pixel values [6, 1].
3.2Principal component analysis
Principal component analysis (PCA) vas developed for purpose of recognizing faces and for compressing images of faces [7, 8]. The method is based on mathematical procedure, which transforms a set of uncorrelated features, also known as principal components, into smaller set of correlated features.
3.3Association rules
Until recently, association rules were not used for texture description. Frequently, they were used for generation of association rules in large databases. For parametrization we used program ARTEX, developed by Bevk [5] and based on fast Apriori algorithm [10, 12].
- Machine learning
After we described images with numerical parameters we processed received data with different methods of machine learning.
The method of machine learning usualy splits all available data into two independent data sets. One data set is used for training the classifier and the other data set is used for testing accuracy of hypothesis. Relation between incorrectly classified cases and all cases is called classification error. We also must consider default error, which tells us how difficult the classification problem is. It is equal to relation between the number of cases in minority class and number of all cases.
In all our cases of ML we used 10-fold cross validation:
-Starting traning set is randomly divided into ten aproximately equal data sets;
-Nine data sets are used for learning classifier and one is used for testing it;
-The whole procedure is repeated ten times and the result is average error of testing the accuracy of classifier.
We used classification error for representing results of ML.
4.1Description of different ML algorithms
We used four different algorithms for machine learning:
-Decision trees [2],
-Naïve Bayesian classifier [14],
-K-nearest neighbour [11],
-Neural networks [15].
Decision trees are included in See5 application from RuleQuest company, also known as algorithm C5.0 and the other three algorithms are included in application WEKA, developed on Waikato university in New Zeland [9].
4.2Results
For each parametrization type we used several different settings. We give average values of machine learning results for main parametrization types.
We noticed that the PCA algorithm is not appropriate for our problem. After all it was developed for face recognition and not for texture description.
If we ignore PCA results we got very interesting results for aluminium oxide suspension. We had 57 pictures of control class and 44 pictures of experimental class - default error was 43,5%. With machine learning we got only approximately 17% of classification error (Neural networks), which tells us, that suspension has some kind of memory (see table 1).
Parametrization type
/ Classification error from machine learning algorithms [%]C5.0 / Naïve B. / K-NN / Neural N.
Statistical descr. / 25,3 / 24,8 / 19,8 / 17,8
PCA / 42,7 / 41,0 / 49,1 / 49,4
Association rules / 27,7 / 24,6 / 28,0 / 16,8
Table 1: Results for aluminium oxide suspension
The results of other three suspensions was not satisfying and we can explain that with the argument, that each picture was very particular. And that is bad for machine learning. For aluminium oxide, all pictures were very similar. Results of other three suspensions are presented in tables 2, 3 and 4.
Parametrization type
/ Classification error from machine learning algorithms [%]C5.0 / Naïve B. / K-NN / Neural N.
Statistical descr. / 40,5 / 37,7 / 33,6 / 39,6
PCA / 46,4 / 46,4 / 55,1 / 46,0
Association rules / 46,7 / 49,7 / 38,9 / 40,1
Table 2: Results for ordinary tap water
Parametrization type
/ Classification error from machine learning algorithms [%]C5.0 / Naïve B. / K-NN / Neural N.
Statistical descr. / 48,8 / 46,7 / 35,7 / 34,7
PCA / 48,0 / 49,3 / 49,7 / 51,7
Association rules / 45,9 / 47,5 / 50,2 / 39,4
Table 3: Results for filtered tap water with inverse osmosis
Parametrization type
/ Classification error from machine learning algorithms [%]C5.0 / Naïve B. / K-NN / Neural N.
Statistical descr. / 38,6 / 41,5 / 31,4 / 31,9
PCA / 46,1 / 49,2 / 48,4 / 43,5
Association rules / 40,3 / 41,7 / 46,5 / *
Table 4: Results for spring water from Tunjice
- Statistical analysis of parameters
Besides machine learning we were also looking for differences between control and experimental class with statistical t-test. With t-test we were making comparison between medians for each parameter and we tryed to find differences. The results were similar to results of machine learning.
- Conclusions
The results of machine learning and statistical tests indicate the same conclusion: Aluminium oxide suspension shows significant differences between experimental and control liquid, while the other three suspensions do not exhibit significant differences.
Among three parametrization techniques statistical parametrization and association rules seems to be able to extract significant information from image textures, while PCA is useless. Among different ML algorithms neural networks seem to perform best.
Our results show that
-Aluminium oxide suspension is a good candidate for testing the memory of water.
-Tap water has too many unknown minerals and is therefore inaproprate for experimentation.
-Water cleaned with reverse osmosis has too low amount of minerals and therefore the eventual information stored in the water cannot be seen.
-Spring water from Tunjice was not even expected to show any differences as it is by default exposed to natural energy source.
Our results with Aluminium oxide suggest that
-Natural energy source from Tunjice has some influence;
-The water indeed has a memory;
-The memory of water is detectable using the dark field microscopy technique together with numerical parametrization of images and machine learning (namely, only by human observation of images we were not able to find the differences).
Further experiments are necessary in order to reconfirm the above conclusions and to verify the repeatability of the results.
References
[1] Haralick R.M., Shanmugam K., Dinstein
I., Textural Features for Image Classi¯ca-
tion, IEEE Transactions on Systems, Man
and Cybernetics, 1973, pages 610-621.
[2] Quinlan J.R., C4.5 programs for machine
learning, Morgan Kaufmann, 1993.
[3] Bevk M.,Kononenko I., Exploring the memory of matter, Proc. 5th Conference on cognitive science, Ljubljana, 2002, pages 115-117.
[4] Kroeplin B., Hein M., Herrmann J.P., Heusel
B., Project APOLLO IV, Report on the Mi-
crooptic Investigations of Water and Watery
Solutions, Institute for Statics and Dynam-
ics of Aerospace Structures, University of
Stuttgart, 2000.
[5] Bevk M., Texture Analysis with Machine Learning, M.Sc. Thesis, University of Ljubljana, Faculty of Computer and Information Science, Ljubljana, Slovenia. 2003 (in Slovene)
[6] Bevk M.,Kononenko I., A Statistical Approach to Texture Description of Medical Images: A Preliminary Study, Workshop on Machine Learning in Computer Vision, Nineteenth International Conference on Machine Learning, Sydney, Australia, 2002, pages 39-48
[7] L. Sirovich and M. Kirby, A low-dimensional procedure for the characterisation of human faces. Journal of the Optical Society of America, 1987, pages 519-524.
[8] M. Turk and A. Pentland, Eigenfaces for recognition. Journal of Cognitive Neuroscience, 1991, pages 71-86.
[9] I. H. Witten, E. Frank, Data mining: Practical machine learning tools and techniques with Java implementations, Morgan Kaufmann, 2000.
[10] R. Agrawal and R. Srikant, Fast algorithms for mining association rules. In J. B. Bocca, M. Jarke, and C. Zaniolo, editors, Proc. 20th Int. Conf. Very Large Data Bases, VLDB, pages 487-499. Morgan Kaufmann, 1994.
[11] Aha, D., and D. Kibler, "Instance-based learning algorithms", Machine Learning, vol.6, pp. 37-66, 1991.
[12] R. Agrawal, T. Imielinski, and A. Swami, Mining association rules between sets of items in large databases. In P. Buneman and S. Jajodia, editors, Proceedings of the 1993 ACM SIGMOD International Conference on Mangement of Data, pages 207-216, Washington, D.C., 1993.
[13] R. Haralick, K. Shanmugam, and I. Dinstein, Textural features for image classification. IEEE Transactions on Systems, Man and Cybernetics, pages 610-621, 1973.
[14] G. H. John and P. Langley, Estimating Continuous Distributions in Bayesian Classifiers. Proceedings of the Eleventh Conference on Uncertainty in Artificial Intelligence. pp. 338-345. Morgan Kaufmann, San Mateo,1995.
[15] D.E. Rumelhart, G.E. Hinton, R.J. Williams, Learning internal representations by error propagation. In: Rumelhart D.E. and McClelland J.L. (eds.) Parallel Distributed Processing, Vol. 1: Foundations. Cambridge: MIT Press, 1986.