Emma Amnemyr and Daniel Björklun present their master’s thesis on Wednesday 29 May at 10:15 in MH:309A.

Abstract:
Arabica coffee production in Africa has declined significantly over the last half century, partly due to Coffee Berry Disease (CBD), caused by the fungus Colletotrichum kahawae. This disease results in substantial economic losses, estimated at USD 350-500 million annually. Recent advancements in machine learning (ML) and computer vision offer powerful tools for disease detection. However, annotating data for training ML models is both time-consuming and costly. Active Learning (AL) aims to maximize annotation efficiency by strategically selecting data points for annotation, thereby accelerating model performance improvement. This thesis evaluates the impact of utilizing both strong and weak labels in AL for detecting CBD. Initially, an AL framework was implemented, and four query strategies using only strong annotations were developed and evaluated. One of these strategies, ALCU Soft-Rank, showed promise and appeared to outperform the baselines. This strategy was then further developed to determine whether the inclusion of weak labels could enhance the performance. The results indicated that, under the chosen conditions, incorporating weak labels was not beneficial, and the original ALCU Soft-Rank utilizing only strong labels performed best. Further exploration of active learning in this setting, especially using other base models, would be interesting.

Examiner: 
Alexandros Sopasakis, Centre for Mathematical Sciences, Lund University

Supervisors:
Kalle Åström, Centre for Mathematical Sciences, Lund University
Aleksis Pirinen, RISE
Olof Mogren, RISE