SBIR-STTR Award

Our objective is to implement neural net technology for the automatic identification of plant specimens. The ultimate focus of our work will be the identification of medicinally-important plants of a particular geographic sub-region. Because of the difficulty of identifying plants based on flowering or fruiting material (usually requiring careful dissection of parts) we will initially use leaves for identification. Leaves are mostly two- dimensional easily removed from a plant or digitized from an herbarium specimen and have a number of recognizable features. Existing methodology, however, relies on rather subjective, human identification. We will utilize a backpropagation neural network to achieve rapid identification of plant specimens based on training the network with a representative of set of leaves. Our goal is to allow the identification (to species or a subset of species) to be made by persons not trained in plant taxonomy. This research could have applications in the identification of rare and endangered species, agricultural weeds, and toxic plants. The latter could be especially useful in allowing the quick identification of plants of a given region that are potentially poisonous or carcinogenic to humans.Proposed commercial application:This research will lead to an improved method for quantitative identification of plant specimens, identification of rare and endangered species, agricultural weeds, and toxic plants. The latter could be especially useful in allowing the quick identification of plants of a given region that are potentially poisonous or carcinogenic to humans.National Cancer Institute (NCI)

LONG-TERM OBJECTIVES 1. Develop a computerized system, based on hierarchical neural network pattern recognition technology, for reliable identification of plants. 2. Identify poisonous plants. 3. Expedite discovery of new medicinal plants. 4. Create an image database directly from plant material and link with existing medicinal plant databases. 5. Develop commercial product for pharmaceutical companies, agriculture and others. SPECIFIC AIMS 1. Design hierarchical system of neural networks to follow natural plant taxonomy groupings and extend our identification technology to a large number of plant species. 2. Improve accuracy of identification. 3. Design a prototype workstation for botanical and agricultural field stations and laboratories. RESEARCH DESIGN AND METHODS FOR ACHIEVING GOALS. 1. Digitize large number of plant species from special collections. 2. Measure automatically venation patterns and shape. 3. Design hierarchical neural networks to divide plants into natural groupings. 4. Accumulate virtual herbarium database as leaves are digitized (scanned or photographed). POTENTIAL FOR TECHNOLOGICAL INNOVATION This system is unique in capturing botanical recognition knowledge in a hierarchy of neural networks and is the first fully-computerized system for plant identification utilizing information digitized directly from plants. PROPOSED COMMERCIAL APPLICATION 1. Expedite discovery of new medicinal plants for pharmaceutical industry. 2. Create valuable database directly from plants. 3. Identification of poisonous plants. 4. Valuable for rapid identification of invasive weeds.

Thesaurus Terms:
artificial intelligence, classification, medicinal plant, plant morphology carcinogen, digital imaging, geographic site, information system, leaf, plant poison, systematic biology