Angel Gutierrez

Visualization of hyperspectral imagery

We developed four new techniques to visualize hyper spectral image data for man-in-the-loop target detection. The methods respectively: (1) display the subsequent bands as a movie (“movie”), (2) map the data onto three channels and display these as a colour image (“colour”), (3) display the correlation between the pixel signatures and a known target signature (“match”) and (4) display the output of a standard anomaly detector (“anomaly”). The movie technique requires no assumptions about the target signature and involves no information loss. The colour technique produces a single image that can be displayed in real-time. A disadvantage of this technique is loss of information. A display of the match between a target signature and pixels and can be interpreted easily and fast, but this technique relies on precise knowledge of the target signature. The anomaly detector signifies pixels with signatures that deviate from the (local) background. We performed a target detection experiment with human observers to determine their relative performance with the four techniques,. The results show that the “match” presentation yields the best performance, followed by “movie” and “anomaly”, while performance with the “colour” presentation was the poorest. Each scheme has its advantages and disadvantages and is more or less suited for real-time and post-hoc processing. The rationale is that the final interpretation is best done by a human observer. In contrast to automatic target recognition systems, the interpretation of hyper spectral imagery by the human visual system is robust to noise and image transformations and requires a minimal number of assumptions (about signature of target and background, target shape etc.) When more knowledge about target and background is available this may be used to help the observer interpreting the data (aided target detection).

The PSO Algorithm and the Diagnosis of Multiple Sclerosis Using Artificial Neural Networks

Artificial neural networks with radial basis functions are used to diagnose patients with multiple sclerosis. But the training of this type of network requires a great amount of time. It would be advantageous if we could speed up this training. The Particle Swarm Optimization (PSO) algorithm was previously used in a particular case, but the results were poor. It was suggested to review this scenario using different network architectures, with other selection criteria for the most significant features of the data, and other training sets. In this paper we follow these ideas and work with various hidden nodes and input nodes. The most significant coefficients were extracted following the Kolmogorov-Smirnov test, the Principal Component Analysis and the Largest Coefficient criteria. The training process used the left-out method. The neural networks were then trained using the Particle Swarm Optimization technique, and the gradient descent algorithm. Once the networks had been trained, we tested them using the original left-out element. We repeated the process, taking averages of all the cases. Our results confirmed that the standard PSO algorithm is not a good method. But then we used the values that gave bad results with the PSO approach as the starting values for a network trained with the gradient descent. In some of the cases, the results were improved when compared to the case of starting with random values.

Hyperspectral image processing: a direct image simplification method

We describe a novel approach to produce color composite images from hyperspectral data using weighted spectra averages. The weighted average is based on a sequence of numbers (weights) selected using pixel value information and interband distance. Separate sequences of weights are generated for each of the three color bands forming the color composite image. Tuning of the weighting parameters and emphasis on different spectral areas allows for emphasis of one or other feature in the image. The produced image is a distinct approach from a regular color composite result, since all the bands provide information to the final result. The algorithm was implemented in high level programming language and provided with a user friendly graphical interface. The current design allows for stand-alone usage or for further modifications into a real time visualization module. Experimental results show that the weighted color composition is an extremely fast visualization tool.

Influence of wavelet boundary conditions on the classification of biological signals

Doctors utilized Brainstem Auditory Evoked Potentials (BAEP) to diagnose patients with multiple sclerosis. We use eight coefficients of each of the several wavelet transforms of the BAEP signals to train an artificial neural network with radial basis functions. We study how the boundary conditions used to determine the wavelet transforms affect the maximum number of correct diagnoses. Using this information, we establish the best strategy to avoid misleading information created by the boundary conditions.