Training a new cotton imaging system via a transfer learning approach

In this paper, a transfer learning case study on cotton quality evaluation is presented whereby a new prototype imaging system (target problem) is trained using knowledge transferred from a reference system (source problem). We describe the properties of both systems and explain how our problem setup is a specific case of inductive transfer learning. We then present a feature-based domain adaptation framework to reduce domain divergence, allowing data to be compared between the two systems. Finally, we discuss ways of transferring domain-specific knowledge from the reference system to the new system. We demonstrate our approach on cotton data available from the reference system, and those that we generate using our prototype system.