Condition-based synthetic dataset for amodal segmentation of occluded cucumbers in agricultural images

• Published in: Computers and electronics in agriculture Vol. 238; p. 110800
• Main Authors: Son, Jin-Ho, Song, Hojun, Song, Chae-yeong, Ha, Minse, Kang, Dabin, Ha, Yu-Shin
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.11.2025
• Subjects: Amodal segmentation; Computer vision; Occlusion; Precision agriculture; Synthetic dataset generation; Occlusion; Precision agriculture; Computer vision; Amodal segmentation; Synthetic dataset generation
• ISSN: 0168-1699
• DOI: 10.1016/j.compag.2025.110800

•Novel framework generates synthetic occlusion datasets for amodal crop segmentation.•Condition-based method simulates realistic leaf occlusion in agricultural settings.•Gamma augmentation enhances model robustness across diverse lighting conditions.•AISFormer achieves accurate cucumber segmentation under complex occlusion scenarios. Occlusion, caused by overlapping leaves and dense foliage, poses significant challenges for crop segmentation in agricultural environments. Traditional segmentation methods struggle to handle these occlusions, particularly in complex and dynamic agricultural settings. As agricultural environments are subject to varying lighting conditions and environmental factors, the ability to detect and segment crops accurately remains a persistent issue in precision farming. This study aims to address these challenges through the development of a Synthetic Dataset Generation Framework that replicates realistic occlusion scenarios using a condition-based approach. Systematic adjustments to leaf composition, position, and occlusion ratios facilitated the framework to generate synthetic datasets representative of the diverse and complex conditions found in agricultural environments. To enhance realism, gamma correction-basedbrightness and contrast augmentations were applied to simulate both low-light and high-light conditions. These augmentations enhanced dataset diversity to better replicate real-world illumination variations. The ability of the model to handle dynamic lighting and complex occlusion scenarios was enhanced through this approach, advancing precision agriculture and contributing to sustainable farming practices by providing more reliable, adaptable segmentation solutions for real-world agricultural applications.