Submitted to: Meeting Abstract
Publication Type: Abstract Only
Publication Acceptance Date: 8/27/2021
Publication Date: 9/1/2021
Citation: Pelletier, M.G., Holt, G.A., Wanjura, J.D. 2021. Virtual cotton field for accelerated deep learning development. Virtual NP-306 Seminar. Oral presentation. September 1, 2021. Bletsville, MD.
Technical Abstract: Plastic contamination is a driving force behind the loss of $750 million U.S. in market value. As such the removal of plastic contamination from cotton is a top priority to the U.S. cotton industry. One of the main sources of plastic contamination is the plastic module wrap from the John Deere round module harvesters. Plastic still finds a way into the cotton that is brought to the processing plant, despite diligent efforts by plant’s personnel to remove all plastic encountered during the unwrapping process of the incoming seed cotton storage modules. A major source of contamination is plastic being picked during the harvest operation. Of critical need is the development of exclusion machinery for the harvest to allow it to avoid this in-field plastic contamination. As ambient lighting in the fields is highly variable, traditional machine-vision algorithms are not effective under these wildly varying conditions. It is hypothesized that the new Deep-Learning Artificial Intelligence, AI models might provide a more robust alternative. The downside to the use of these AI models is that they require 100,000's of annotated images, where the annotation requires that inside every image each object of interest is identified with a bounding box. This annotation requirement, plus the extensive number of images, combines to make this approach a costly endeavor if using real-world images. In order to mitigate this hurdle, the research is developing a virtual 3D cotton field where the software can control the lighting, cotton plants, weeds and plastic targets and camera placement. The virtual world then allows for executing experiments that are designed to cover a far greater number of variables than is possible with real-world experiments, at a fraction of the cost as the image annotations are provided by the software directly.