BCMask: a finer leaf instance segmentation with bilayer convolution mask

Whether in natural scenes or laboratory environments, leaf instance segmentation is still a challenging task in high-throughput plant phenotypic research. Because compared with normal instance objects, leaves have more complex boundaries and severe inter-leaf occlusions. In this paper, we present an effective two-stage method called Bilayer Convolution Mask (BCMask) for high-quality leaf instance segmentation. BCMask consists of three main modules: (1) Bottom-up Path Augmentation (BPA) module is added after Feature Pyramid Network (FPN) in Faster R-CNN. BPA shortens the information path between lower layers and high-level layers, and helps accurate semantical features in lower layers to enhance the entire feature hierarchy; (2) Bilayer Occlusion Module. This module consists of two convolutional layers with a residual structure, which decouples the occluding leaves and the partially occluded target leaf during the mask regression; (3) Mask Refining Module. This module uses an iterative refinement method with adaptive selection to classify pixels, which effectively alleviates the problem of inaccurate leaf boundary segmentation. To validate BCMask, this paper takes the chrysanthemum seedling leaf dataset for experiment, which is collected in the natural environment with complex boundaries and severe occlusions. Two remarkable public datasets CVPPA and Komatsuna under laboratory environments are also added as supplements to validate the robustness of BCMask. The proposed method achieves the 60.42% average precision (AP) score outperforming state-of-the-art methods.

Automated summary

We propose a high-quality leaf instance segmentation method called BCMask, which effectively addresses the challenges of complex boundaries and occlusions among leaves. By utilizing the Bottom-up Path Augmentation module, Bilayer Occlusion Module, and Mask Refining Module, BCMask achieves accurate leaf instance segmentation. Experimental results on a chrysanthemum seedling leaf dataset collected in natural environments, as well as two public datasets (CVPPA and Komatsuna) in laboratory environments, demonstrate that BCMask outperforms state-of-the-art methods with an average precision (AP) score of 60.42%.