- Author
- Neal, J. A. | Land, C. E. | Avent, R. R. | Churchill, R. J.
- Title
- Application of Artificial Neural Netwoks to Machine Vision Flame Detection. Final Report. May 1990-November 1990.
- Coporate
- American Research Corporation of Virginia, Radford, VA
- Sponsor
- Air Force Engineering and Services Center, Tyndall AFB, FL
- Report
- ESL-TR-90-49, April 1991, 58 p.
- Distribution
- Available from National Technical Information Service
- Keywords
- flame detection | fire detectors | methodology | neural networks
- Identifiers
- machine vision techniques; neural network computation; visible spectrum signatures of fire; digitized video imagery
- Abstract
- The U. S. Air Force has identified a need for rapid, accurate and reliable detection and classification of fires. To address this need, a proof-of-concept neural network-based, intelligent machine vision interface for the detection of flame signatures in the visible spectrum has been developed. The objective of the work conducted under this Phase 1 program has been to determine the feasibility of using machine vision techniques and neural network computation to detect and classify visible spectrum signatures of fire in the presence of complex background imagery. Standard fire detectors which rely on heat or smoke sensing devices tend to be slow and to react only after the fire reaches a significant level. Current electromagnetic sensing techniques have the desired speed but lack accuracy. The Phase 1 program approach to these problems used machine vision techniques to generate digitally filtered HSI (Hue, Saturation, Intensity)-formatted video data. Once filtered, these data were then presented to an artifical neural network for analysis. In the Phase 1 program, positive results were achieved in the application of neural networks in an intelligent HSI video data classification and analysis system for the detection of fires. The principal result of the Phase 1 effort was the implementaion of a proof-of-concept fire detection system. Additional results included the development of image processing modules capable of intensity and hue thresholding, low pass filtering, image subtraction, region detection and labeling and HSI data normalization. In the Phase 1 system, these image processing modules were used to filter and format image data for processing by a neural network. Work conducted during the Phase 1 program resulted in a highly accurate neural network architecture. The Phase 1 neural network was trained to recognize expanding fire regions within an image using 137 training data sets consisting of 96 fire region sets and 41 false alarm region sets. After training was completed, this network was presented 23 test data sets containing 17 fire regions and 6 false alarm regions. The network demonstrated 100 percent accuracy with the training data sets and was also 100 percent accurate with the test data sets. This system was able to demonstrate reliable and repeatable detection of fire regions scaled to 4 by 4 feet at a range of 150 feet. The potential applications for this system, once fully developed in Phase 2 of the program, include installation in facilities requiring fire detection systems. For the U. S. Air Force, this would include aircraft hangars, ammunition depots and any facility containing high value assets and flammable materials. Phase 2 of the program will accomplish the feasibility established in converting the system module algorithms to hardware implementations, thus significantly increasing system processing speed and reducing fire detection times to meet ever more demanding U. S. Air Force specifications.