LV-SEGFORMER: Towards More Accurate Leaf-Vein Segmentation with Transformer

Abstract

Segmenting vein from leaf images is challenging because of the strong expansibility, high connectivity and irregular distribution of vein patterns. Although several recent attempts have been made to apply CNNs to leaf vein segmentation, they are weak to capture global contexts of lower-order veins and fine-grained features about higher-order veins due to the continuous stacking of convolution and pooling operations. Inspired by the great successes of transformers on semantic image segmentation, we for the first time develop a novel transformer based model, named LV-SegFormer for leaf vein segmentation. The LV-SegFormer leverages the powerful global modeling capability of transformer to generate advanced representations. The multi-stage alignment module and the hybrid decoder are designed to incorporate contexts from different stages and recover fine-grained features for performance boosting. Extensive experiments are conducted on two challenging benchmark datasets and the results demonstrate that the proposed LV-SegFormer achieves higher performance over the state-of-the-arts.