Análise de técnicas de segmentação de ervas daninhas em lavouras usando imagens de drones

Abstract

Currently, semantic segmentation techniques have been crucial in computer vision tasks. This approach aims to assign labels to each pixel in an image, enabling a deeper understanding of the scene. Over the years, several architectures have been proposed, with neural networks such as U-Net, DeepLabV3+, and FPN standing out among others. In the context of this work, two datasets widely used in the field of precision agriculture were employed: Cofly and AgricultureVision, both composed of aerial images of agricultural areas captured by UAVs. The main objective is the segmentation of weeds among crops. The experimental methodology was structured into two main stages. In the first, several combinations of segmentation architectures and different backbones were evaluated in order to identify the most promising configurations. The initial results showed that the U-Net architecture, combined with the VGG19 backbone, achieved an IoU of 0.3833 on the Cofly dataset. Similarly, the FPN architecture integrated with the DenseNet169 backbone reached an IoU of 0.3779 on the Agriculture-Vision dataset. Although the observed performances indicate potential, comparative analyses with studies from the literature suggest that these results could be further improved. Based on this premise, an ensemble strategy was adopted, aiming to combine different models to increase segmentation robustness and accuracy. This approach proved effective in certain scenarios. By employing different combinations of the FPN architecture, it was possible to achieve superior performance, reaching a score of 0.4005 — the best result obtained in this study for the Agriculture-Vision dataset — and 0.3833 for Cofly.