An instance segmentation framework based on parallelogram mask for crop row detection in various farmlands

Crop row detection is one of the essential steps for autonomous guidance in agriculture. Conventional methods only detect the center lines of crop rows, without providing information about their widths and shapes, which cannot meet the growing demands. Instance segmentation, which segments each object with an individual pixel-wise mask, seems to be a more appropriate solution. However, universal instance segmentation methods usually detect with noise masks belonging to other crops or weeds. To address this issue, we propose a customized instance segmentation framework consisting of two steps. First, an adaptive deep neural network transforms the image into an approximate aerial view, in which the crop rows resemble parallelograms. Subsequently, we propose an instance segmentation approach called Parallelogram Mask (PlgMask) to segment the crop rows within the transformed image. We train and evaluate our method on the CRBD dataset Vidovíc et al. (Pattern Recognit 55:68–86, 2016), and the results show that it can accurately detect crop rows without noise masks. Additionally, we evaluate our method under the zero-shot setting, which demonstrates that the proposed method can achieve great performance even on an unseen dataset.