Identificação de mamonas em plantações de cana-de-açúcar utilizando aprendizado ativo

Abstract

Castor bean plants have applications in multiple industries but can become a weed in several crops. Manual identification and control are unfeasible in large plantations. Using machine learning models can make the identification process efficient and accurate. Still, it is necessary to label a large amount of data, which is laborious and inefficient since many of the images are similar and do not improve model training. Active learning allows labeling only data that positively impacts model performance, which can match or surpass the performance of a model trained on the entire set, with significantly less data labeling effort. In this work, we test an active learning method that uses self-supervised pretext tasks to separate the data that must be labeled. The models trained with the pretext tasks presented a decrease in the recall relative to the reference model, which has 68%. We also trained a pseudo task that creates batches with the ideal distribution. The pseudo task's classifier obtained a recall of 86%, while being trained in less than 1% of the dataset.