Structural Analysis of Asian and African Rice Panicles via Transfer Learning

Rice panicle architecture affects yield and grain quality in both Asian (Oryza sativa) and African rice (Oryza glaberrima). Manual low-throughput phenotyping methods are inaccurate and non-repeatable. This study introduces a novel image-based method for automated junction detection using transfer learning and advanced image processing, leveraging visual similarities between pavement cracks and rice panicles. Seven pre-trained pixel-level crack segmentation models are employed on the panicle images to extract their morphology. Skeletonization and some image processing techniques are then applied to detect axis junctions from the resulting binary images. Experiment results revealed U-Net’s robust performance with nearly 90 % accuracy in segmentation tasks. Zhang-Suen thinning combined with Crossing Number reached about 80 % accuracy for main axis junctions detection. High-order junction detection was challenging, especially for the complex Asian panicles. This study shows a correlation between pavement cracks and rice panicles, demonstrating the potential for automated rice phenotyping. Future efforts will focus on enhancing detection accuracy and extending the approach to include grain detection.