MultiTaskVenationNet: A Multi-Task Deep Neural Network with Strip Pooling and Hybrid Upsampling for Leaf Vein Segmentation

Leaf vein segmentation is a critical task in plant phenotyping and species classification, yet it remains challenging due to the hierarchical, curvilinear nature of veins and interference from complex backgrounds. Existing methods face three key limitations. First, they lack directional context modeling, leading to blurred vein boundaries and the omission of fine venation. Second, they fail to effectively capture global dependencies, limiting semantic coherence across spatial regions. Third, they do not incorporate explicit mechanisms for detecting vein discontinuities, which is essential for complete topological understanding. To address these challenges, we propose MultiTaskVenationNet (MTV-Net), a multi-task deep segmentation framework that integrates four complementary modules. The Strip Pooling Module (SPM) captures orientation-specific long-range context by performing directional pooling along horizontal and vertical axes, enhancing the visibility of delicate vein structures. The Global Context Block (GCBlock) aggregates long-range dependencies through channel attention at the bottleneck stage, improving the semantic consistency of encoded features. A dual-branch decoder explicitly separates the learning objectives for vein segmentation and breakpoint detection. At the same time, a hybrid upsampling strategy combines bilinear interpolation and transposed convolution to accurately reconstruct vein boundaries without introducing artifacts. Extensive experiments on the LVD2021 benchmark dataset demonstrate that MTV-Net outperforms state-of-the-art models such as U-Net, GCNet, CE-Net, and HRNet, achieving an IoU of 76.46 ± 0.27 and a Dice coefficient of 86.61 ± 0.18. The model also exhibits strong generalization across diverse leaf morphologies, vein densities, and lighting conditions, validating its effectiveness and robustness for high-precision leaf vein analysis.