CGA-ASNet: an RGB-D amodal segmentation network for restoring occluded tomato regions

Obtaining the complete morphology of tomato fruits under non-destructive conditions is essential for phenotype research, yet fruit occlusions often hinder deep learning-based image segmentation methods from capturing the true shape of occluded regions. This limitation reduces prediction accuracy and adversely impacts phenotype data acquisition. To overcome this challenge, we propose CGA-ASNet, an RGB-D amodal segmentation network incorporating a Contextual and Global Attention (CGA) module. A synthetic tomato dataset (Tomato-sim) was constructed using NVIDIA Isaac Sim's Replicator Composer (ISRC) to realistically simulate tomato morphology and greenhouse environments, and the network was trained on this dataset. To evaluate generalization, CGA-ASNet was tested on both the synthetic and a separate real-world dataset. While no explicit domain adaptation techniques were adopted, diverse lighting conditions (strong, normal, and weak illumination) were simulated to implicitly reduce the domain gap, and a mean coordinate fusion algorithm was introduced to improve annotation completeness in real-world occlusion scenarios. By leveraging contextual information among feature input keys for self-attention learning, capturing global information, and expanding the receptive field, CGA-ASNet enhanced representation capacity, semantic understanding, and localization accuracy. Experimental results demonstrated that CGA-ASNet achieved an F@0.75 score of 94.2 and a mean Intersection over Union (mIoU) of 82.4% in greenhouse amodal segmentation tasks. These findings indicate that training with well-designed synthetic datasets can effectively support accurate occlusion-aware segmentation in real environments, providing a practical solution for tomato phenotyping in greenhouse conditions.