Neural Networks for Computer Vision
 
Investigator:  Dr. Suchendra M. Bhandarkar
 
    This project investigates the applications of neural networks to problems in computer vision. The neural network currently under investigation is the Hierarchical Self Organizing Map (HSOM). The HSOM has been shown to generate a multiscale or hierarchical segmentation of gray scale, range and multispectral images. Image segmentation is a process of partitioning an input image into disjoint sub regions which individually satisfy the properties of homogeneity and connectivity. Multiscale structures and algorithms that unify the treatment of local and global scene information are of particular importance in image segmentation. The complexity analysis of most biological (in particular, human) visual systems strongly indicates that hierarchical or multiscale internal representations and processing techniques are employed during scene interpretation. From the computer scientist's point of view, hierarchical or multiscale representations and processing techniques offer a means of integrating local and global scene information and also alleviating the combinatorial complexity of scene interpretation.
 
    Furthermore, in most computer vision tasks, the appropriate choice of scale is a crucial issue. In most images, the objects appear with varying sizes, at varying positions and orientations and at varying degrees of spatial resolution.  Consequently, there often does not exist a single scale of spatial resolution that could be deemed appropriate for analysis of the entire image.  A choice of a single scale may prove too fine in certain portions of the image resulting in unnecessary detail, noise and clutter and too coarse in other portions of the image resulting in loss of desired detail. Clearly, a multiscale analysis of the image is necessary which naturally entails a segmentation technique that is capable of generating a multiscale representation of the image data.
 
 
    A severe shortcoming of the traditional single layer self organizing map (SOM) is that the number of neural units in the competitive layer needs to be approximately equal to the number of regions desired in the segmented image. However, it is usually not possible to determine a priori the correct number of regions in the segmented image. This severely limits the usefulness of the conventional single layer SOM for image segmentation. The research thus far has resulted in the formulation and implementation of a hierarchical SOM (HSOM) and the learning procedure that makes it suitable for image segmentation. The HSOM combines the ideas of self organization and topographic mapping with those of multiscale image segmentation, thus making it distinctly unique from previous approaches to image segmentation using neural networks. The HSOM generates an "Abstraction Tree" which represents the segmented image at multiple scales of resolution. Proper traversal of the abstraction tree results in the final segmented image. The performance of the HSOM has been demonstrated on intensity and range images. Current efforts are directed at showing the usefulness of the HSOM for segmentation of multispectral MRI images. Future research will be devoted to learning image segmentation criteria from examples which in our case amounts to learning of traversal strategies for the abstraction tree.
 
 
References
 
S.M. Bhandarkar, J. Koh and M. Suk, Multi-scale Image Segmentation using a Hierarchical Self-Organizing Feature Map, Neurocomputing, Vol. 14, No. 3, 1997, pp. 241-272.
 
S.M. Bhandarkar, J. Koh and M. Suk, A Hierarchical Neural Network and Its Application to Image Segmentation, invited paper in  Intl. Journal on Mathematics and Computers in Simulation, Vol. 41, Nos. 3-4, 1996, pp. 337-355.
 
J. Koh, M. Suk and S. M. Bhandarkar, A Multi-layer Self-Organizing Feature Map For Range Image Segmentation, Neural Networks, Vol. 8, No. 1, 1995, pp.  67-86.
 
S.M. Bhandarkar, J. Koh and M. Suk, Hierarchical Self-Organizing Neural Networks and their Applications to Image Processing, invited paper in Proc. 14th IMACS World Congress on Computational and Applied Mathematics, Atlanta, GA, July 1994,  Vol. 2, pp. 573-576.
 
J. Koh, M. Suk and S. M. Bhandarkar, A Multi-layer Kohonen's Self-Organizing Feature Map For Range Image Segmentation, Proc. Intl. Conf. on Neural Networks, San Francisco, CA, April 1993, pp. 1270-1275.
 
J. Koh, M. Suk and S.M. Bhandarkar, A Self-Organizing Neural Network for Hierarchical Range Image Segmentation, Proc. IEEE Intl. Conf. on Robotics and Automation, Atlanta, GA, May 1993, pp. 758-763.