Lighting independent detection of in-field plastic contamination

Authors: Pelletier, Mathew; Wanjura, John; Holt, Gregory

Submitted to: ASABE Annual International Meeting
Publication Type: Abstract Only
Publication Acceptance Date: 7/12/2021
Publication Date: 7/12/2021
Citation: Pelletier, M.G., Wanjura, J.D., Holt, G.A. 2021. Lighting independent detection of in-field plastic contamination. ASABE Annual International Meeting.

Interpretive Summary: Plastic contamination is a driving force behind the loss of $750 million U.S. in market value for cotton. As such the removal of plastic contamination from cotton is a top priority to the U.S. cotton industry. This presentation covers the development of a virtual cotton field simulation that is designed to lower the cost and streamline the development of deep-learning based plastic detection algorithms. These new algorithms are being developed for use in robotic detection of plastic in cotton fields under wide range of illumination (natural ambient lighting) and for difficult to detect plastic colors inside cotton gins. The report covers the status of the virtual environment development that will be utilized to lower the cost and speed development of the deep-learning models.

Technical Abstract: The most cost effective machine-vision systems are based on low-cost color cameras. The traditional machine-learning classifiers in use rely predominantly on color differences for detection of plastic. While this works for the vast majority of plastics brought into the cotton gins; it still leaves about 10-20% of the plastic contamination undetected as the colors are too close to normal cotton colors to allow for detection without significant risk of false-positives. For in-field plastics, the number of hard to detect plastics colors is greater due to wider range of colors presented by background as modified by the variation in natural lighting, where color temperature of the sky-light can vary from 3000k to over 10,000k, Hence a primary research interest, for removal of plastic contamination from field and in cotton gins, is the identification of colored plastics where the plastic color is very similar to the normal background colors. This research development effort seeks to develop alternative classification methods, such as the use of advanced deep-learning methods for the detection of these hard to detect plastics. These deep learning models require 100,000's of images in order to provide accurate results, which creates a steep obstacle before they can be applied to the problem. This research update covers the development of a virtual simulation environment designed to automatically generate a synthetic image data-sets necessary for the design of deep-learning models, thereby lowering the cost and burden in their development.