Metodologia de classificação e segmentação semântica de plantas daninhas com auxílio de conjuntos de dados sintéticos

Abstract

The agricultural sector plays a crucial role in food security and global economic growth, constantly requiring technological advancements to optimize the cultivation process. A significant challenge faced in crop fields is the presence of weeds, which compete with crops for vital resources such as nutrients, water, and light, potentially causing serious damage to the harvest if not identified and treated promptly. In this context, this work proposes the development of a Deep Learning algorithm aimed at identifying and classifying the main species of weeds found in western Paraná. To achieve this goal, advanced techniques in computer vision and image processing will be employed, such as semantic segmentation, mathematical morphology, optimizers, and cost functions. For the classification task, a dataset containing the main weeds present in western Paraná was used, achieving an accuracy of 0.98. An innovative approach of this project is the creation of synthetic datasets, which are fundamental for a more robust and efficient training of the algorithm, allowing it to recognize a wide variety of scenarios and conditions found in Brazilian crop fields. These synthetic datasets will be used in the task of semantic segmentation, replacing the step of acquiring images of weed infestations in maize crops. The chosen evaluation metric for this task was the MioU (Mean Intersection over Union), which measures the model’s accuracy at the pixel level. The model achieved a result of 0.7698.