YOLOv12-RSLW: An Efficient Detection and Severity Grading Framework for Rice False Smut via Count-Area Calibration

Rice false smut is a major panicle disease that affects rice yield and grain quality and is an important target for resistance evaluation in breeding programs. Accurate field phenotyping is important for disease assessment and resistance screening, yet current assessment relies heavily on manual visual scoring and smut ball counting, which are laborious and subject to evaluator variation. In close-range single-panicle images, false smut balls are often small, dense, occluded, adhered, making automatic detection and severity grading difficult. To address these challenges, we developed a YOLOv12-RSLW detector by integrating RepGhost, SimAM, LSCD, and WIoU into YOLOv12. Detection boxes were then used to guide the Segment Anything Model for panicle and lesion mask extraction, allowing calculation of the lesion-to-panicle area ratio as a supplementary indicator for count-based severity grading. A total of 1911 original field images were collected. After augmentation, the dataset contained 5663 images, including 4531 training images, 566 validation images, and 566 test images. Detection performance was evaluated on the test set, while SAM segmentation was assessed using 80 manually annotated original images. YOLOv12-RSLW achieved 92.06% mAP@0.5, 91.46% precision, and 87.01% recall, with 3.45 M parameters and 6.0 GFLOPs. Compared with the baseline YOLOv12, mAP@0.5 and recall increased by 3.60 and 4.37 percentage points, respectively. Within the augmented dataset, 41.2% of samples initially assigned to Grade 1 and 28.1% of those assigned to Grade 2 met the area-ratio criteria for potential reassignment to higher grades. The framework provides a quantitative approach to rice false smut severity phenotyping and may support future resistance breeding after further validation.