Author: Theodor Emanuelsson 

University: Linköping University

Sanitizing equipment and premises is crucial in various businesses, especially in food production to ensure the production of clean and safe food by preventing cross-contamination. Cross-contamination occurs when microorganisms unintentionally transfer from contaminated surfaces to other surfaces during food handling. 

Robotic cleaning systems with sensors and machine learning algorithms can detect and clean potential areas of contamination to ensure thorough disinfection. The use of computer vision-based approaches can detect and classify debris and contaminants, guiding the robot towards relevant areas or validating the cleaning procedure. However, automation for sanitation purposes is underutilized and relies heavily on the experience of workers. 

This thesis explores the use of a novel multispectral imaging system, coupled with deep learning methods, to detect visible organic residue on surfaces commonly found in food processing plants. This imaging system captures additional light outside of the visible spectrum and can detect objects that are invisible to the human eye, providing useful information for a detection system.

This thesis specifically investigates which type of source dataset is the most useful for transfer learning, the impact of different spectral bands on the solution, and whether multispectral-specific methods are feasible with limited data.