A Transfer Learning Network Based on an Attention Mechanism to Overcome Morphological Differences Across Plant Organs

Plant organ segmentation supports image-based phenotyping, but occlusion, slender structures, and differences between image domains limit performance. This study proposes MVA-PlantNet, an encoder–decoder network incorporating a Multi-view Attention (MVA) block. MVA combines Dynamic Group Depth Attention (DDA), which models channel and within-batch relationships, with Global Spatial Attention (GSA), which uses feature gating. The network was pretrained on synthetic images from the Sweet Pepper Dataset and fine-tuned on real Sweet Pepper images and a Capsicum Dataset containing 30 images, of which six were used for fine-tuning and 24 for validation. Additional fine-tuning experiments used 100 images from the Plant Organ Dataset. MVA-PlantNet achieved a reported mean intersection over union (mIoU) of 71.45% on synthetic Sweet Pepper images, exceeding U-Net and SegFormer by 9.60 and 24.91 percentage points, respectively. The reported mIoUs after fine-tuning were 60.17% on real Sweet Pepper images, 74.16% on the Capsicum Dataset, and 55.52% on the Plant Organ Dataset. MVA-PlantNet had the highest reported mIoU among the evaluated models on these splits. However, stem IoU on the Plant Organ Dataset remained low at 1.32%. The experimental results demonstrate that MVA-PlantNet achieved the best overall segmentation performance among the evaluated models. This showed advantages in segmenting certain small plant organs and adapting to target domains through fine-tuning, providing a feasible approach to addressing segmentation challenges arising from morphological differences among plant organs.