Weed identification in broomcorn millet field using segformer semantic segmentation based on multiple loss functions

Abstract

Computer vision and deep learning are one of the main technologies for weed intelligent recognition in farmland. However, when using the deep learning technology to identify weeds in the field broomcorn millet growing at seedling stage, the weeds grow densely, which not only leads to the imbalance of positive and negative sample data, but also causes the problem of small area segmentation. In this paper, the combination of dice and focal loss was applied to Segformer semantic segmentation network with MiT-B3 as encoder block to identify intensive weeds in seedling broomcorn millet fields. The accuracy scores on the testing set were 95.23 %. The results show that the method proposed in this paper can effectively identify the seedling intensive weeds.