A machine vision system to estimate cotton fiber maturity from longitudinal view using a transfer learning approach

Abstract

This paper describes a proposed machine vision system developed to acquire longitudinal images of complete cotton fibers and then estimate their average maturity using image and pattern analysis. Maturity is important to the cotton industry because it relates to fiber's propensity to break when submitted to mechanical stress and it influences the quality of the goods produced from it (yarns and fabrics). The proposed system is novel because it estimates maturity indirectly from fibers' longitudinal views using auxiliary training data generated from fibers' cross-sectional views. It uses the transfer learning framework to reconcile the distribution differences between the two views before traditional machine learning algorithms are applied to learn a suitable model for cotton fiber maturity imaged longitudinally. In addition, the proposed system is more descriptive than commercially available systems currently used in the cotton industry because it estimates not only the average maturity of a complete cotton fiber, but also the maturity variations along the fiber from end to end. Validation studies performed show that the transfer learning approach is a practical and promising way to train our system.