A high-precision segmentation method for rubber tree stone cells

Stone cells constitute a significant portion of rubber tree bark and are associated with key traits, including bark cracking, hardness, stress resistance, and latex yield. Lack of a fast and accurate method to identify stone cells in rubber tree bark and further for quantifying distribution and area proportion restricts the study of stone cells in the bark of the rubber tree. We propose an automatic segmentation network for rubber tree stone cells based on image recognition, termed CGWO-LWNet. This network addresses challenges such as complex edges, regional distribution patterns, and the instability of traditional segmentation networks during training. Firstly, we introduce a low-rank KAN module to reshape neural network learning, facilitating information sharing and feature fusion between encoders, improving edge segmentation accuracy. Secondly, we design a wavelet attention mechanism, Wave-SC, to capture the distribution patterns of stone cells in rubber tree bark slices. Finally, we propose a new gray wolf constrained optimization algorithm (CGWO) to enhance network training stability. To optimally train CGWO-LWNet, we constructed a dataset of 1084 rubber tree stone cell images from CATAS and conducted experiments. Experimental results show that CGWO-LWNet achieves 69.1% MIoU, 81.7% DSC coefficient, and 80.4% recall on the dataset. Compared to other algorithms, CGWO-LWNet demonstrates higher accuracy, achieving 97.8% in rubber tree bark stone cell segmentation. Our approach offers a practical and robust tool for high-precision segmentation of stone cells, enabling large-scale, accurate trait analysis and facilitating further genetic studies on their development and influence on latex yield, bark integrity, and stress resilience.