Enhancing site-specific weed detection using deep learning transformer architectures

Abstract

Detection of weeds is essential to implement an intelligent weed control system in natural corn fields. Then, to address this issue, the Swin-UNet, Segmenter, and SegFormer deep learning transformer architectures have been implemented and compared. Furthermore, a simple thresholding method has been performed to enhance the segmentation. Moreover, a large pixel-level annotated image dataset acquired under natural field conditions is introduced to train the models. In addition, the well-known Precision, Dice Similarity Coefficient (DSC), Intersection over Union (IoU), and mean Intersection over Union (mIoU) metrics have been used to evaluate the implemented models’ performance. According to the experimental results, the SegFormer architecture was the best model on each of the three proposed weed detection approaches, achieving a macro performance of up to 94.49%, 95.30%, and 91.26% for Precision, DSC, and mIoU, respectively.