Location: Cotton Production and Processing Research
Project Number: 3096-21410-009-016-R
Project Type: Reimbursable Cooperative Agreement
Start Date: Jan 1, 2023
End Date: Dec 31, 2023
Objective:
The global objective of the proposed research is a continuation, of the previous research cycle, which is to assess the best method for detecting plastic contamination at the gin-stand feeder apron and in the field. The specific objectives are to develop adaptive translation tracking algorithms that are immune to variations in ambient lighting and able to detect colored plastics that have the same colors as cotton constituents; such as white plastic against a background of white cotton and black plastic against grey cotton (that is shadowed cotton which shows up in the same region of the “a-b” color plane (L*a*b* color space) as white cotton when shadowed by other cotton bolls and cotton flower-bracts. The Primary Objectives in this year’s research will be to test the efficacy of several identified lighting independent algorithms, that were developed this past year, against the extensive image dataset that was collected by the authors in this year's research effort (in excess of 2.0 Terra-bytes of image data).
Approach:
The incumbents have developed several promising classification algorithms that are immune to variation in lighting differences. The developed algorithms will be tested against an extensive image dataset comprising 2.0 Terra-bytes of image data collected on the gin-stand feeder apron at a rate of 1Hz at two commercial gins using 14 cameras mounted on the gin-stand. From this data, 30,000 images have been curated as being true-plastic-images (TPI). The TPI portion of the image dataset will be augmented using standard Deep-Learning practices to amplify this into 100,000 TPI images. These images will be split into 3 groups, (30, 60, 10)%, that will be used for: {training, testing, validation} of the deep learning models. As these images were collected under real-world commercial conditions in an industrial environment of seed cotton that spans across an entire harvest season. It is anticipated that it should provide a robust dataset to provide a sound basis for ensuring a transferable deep-learning model suitable for deployment into real-world situations. Once the models have been developed and validated; they will be tuned for reduction and conversion into a “Tensorflow lite” model that will be suitable for running on embedded devices. Time permitting, the models will then be ported to run on a NVidia Jetson (Xavier nx) platform for testing of performance efficacy and real-time performance.
