Using Morphological Neural Networks for Robotic Vision Object Recognition

Students: Donna S. Haun and Kristi M. Hummel
February-May, 2000
Dr. Marjorie Skubic, Advisor

This project is based on the OMNN work done by Yonggwan Won and his PhD thesis, "Nonlinear Correlation Filter and Morphology Neural Networks for Image Pattern and Automatic Target Recognition", August 1995. The goal of this project was to detect a target in varied environments and at varying angles. For the project, the training network has one feature map layer and one hidden feed-forward layer, and one output layer which produces the detection image plane (DIP). The training set consisted of 32 images, with the target on light and dark backgrounds with varying degrees of clutter and occlusion. Four different networks were trained using these 32 images, varying the size of the structuring elements and the number of feature maps (7x7 sub element with 2 feature maps, 7x7 sub element with 4 feature maps, 5x5 sub element with 4 feature maps, 3x3 sub element with 4 feature maps). Four different images, with slight foreground clutter, were scanned for target recognition using the above four networks. Evaluation was done by viewing a superposition of the scan results on the original images. Training and scanning times were also measured.
The preliminary conclusions are: -- a 7x7 structuring element with 4 feature maps produces the best results, but has the longest training time; --the 7x7 structuring element with 2 features maps had less good results, but the briefest scanning time; -- a 3x3 structuring element is insufficient for a target this small, but has the briefest training time;

Pictures

The target: a small robot. A few views are shown.

Sample results: 4 different pictures run using the 7x7 element with 4 feature maps

Results showing the 7x7 element with 2 feature maps

Results showing the 5x5 element with 4 feature maps

Results showing the 3x3 element with 4 feature maps

Home

skubic@cecs.missouri.edu
May, 2000