Abstract: Crop kernel phenotyping is time-consuming, labor intensity, and prone to subjective measurement error at present. Most existing seed phenotyping systems can also suffer from kernel adhesion under static imaging. It is often required for the high adaptability over multiple crop species. In this study, an intelligent phenotyping system of crop kernel was developed using waterfall-type line-scan imaging. Four major cereal and legume crops were taken as maize, rice, wheat, and beans. Three functional units were utilized in the system. A dual-stage differential vibratory feeder physically separated kernels using speed-gradient vibration and gravity-driven free fall. Waterfall-like kernel stream was then uniformly distributed before imaging. A monochrome line-scan camera was equipped with a 50 mm fixed-focus industrial lens and operated with transmissive backlight illumination at a spatial resolution of 0.05 mm per pixel and an exposure time of 20 ms. The images were captured at a resolution of 4 096 by 6 000 pixels per frame. Control and weighing unit were coordinated with vibration timing, gate servo actuation, and batch weight acquisition for thousand-kernel weight calculation. Four independent lightweight You Only Look Once version 11 nano (YOLO11n) object detection models were trained for the four crop types on 400 images per crop with 20 000 to 30 000 kernel instances, particularly with 100 fully independent test images for per crop. Sub-pixel edge extraction was performed by fitting B-spline curve to binary foreground contours after Otsu thresholding. An adaptive morphological algorithm selected kernel length from either the minimum bounding rectangle long axis or the maximum Feret diameter, with a shape-asymmetry ratio threshold of 0.95. Thereby the measurement bias was corrected from morphologically irregular kernels, such as curved rice grains and wedge-shaped maize kernels. YOLO11n detection models were validated on the reserved test sets. The results showed that the mean average precision at an intersection-over-union threshold of 0.5 exceeded 97.7% for all four crop models, with precision of 94.4% and recall rate of 96.7%, fully meeting the subsequent requirements for edge extraction and overlap discrimination. In single-kernel morphological accuracy, 50 individually numbered kernels were measured over four varieties of the crop. The coefficient of determination between system and manually measured values exceeded 0.97 for both kernel length and width, with root mean square errors below 0.19 mm and mean absolute errors below 0.15 mm. The sub-pixel extraction reproduced true two-dimensional projection morphology with high fidelity. Batch phenotyping was validated over 40 varieties with 50 independent counting batches per crop. The coefficient of determination between system and manually counted total numbers of crop kernel reached 0.99 over four crops, with mean absolute percentage errors below 0.8% and root mean square errors below 2.0 kernels. Reliable counting performance was achieved over batch sizes from 10 to over 1000 kernels. In the extraction of morphological traits, the coefficients of determination exceeded 0.98 and 0.94, respectively, for the kernel length and width over all tested varieties. Mean absolute percentage errors for kernel length were 2.02% for maize, 2.16% for rice, 1.96% for wheat, and 2.87% for beans; Those values for kernel width were 1.16%, 2.34%, 2.68%, and 3.33%, respectively, all remaining below 3.33%. System stability was then assessed, where Shewhart control charts were applied into 27 consecutive independent runs on an 800-kernel mung bean batch under constant parameters. Kernel length, kernel width, projected area, and projected perimeter remained within the three-sigma control limits, with the coefficients of variation of 0.56% and 0.54% for kernel length and kernel width, respectively. Statistical process control was verified during continuous operation. A two-factor full-factorial experiment was conducted with four stages of feeding cycle (7, 5, 3, and 1 s) and three levels of vibrator driving voltage (69, 67, and 65 V) over 36 replicated runs. The feeding cycle was dominated in the influencing factor on counting net error, with the squares of 1 660 697.64, compared with 52 702.39 for driving voltage, both statistically significant at P less than 0.001. The optimal combination of parameters was obtained with a 5-second feeding cycle and 69 V driving voltage. The better performance of the system was achieved to process 1,950 wheat kernels in an average of 82.7 s with a counting accuracy of 99.4%, and a throughput of 1 415 kernels per minute, compared with the conventional manual counting, commercial seed counters, and tray systems. The waterfall-type line-scan imaging resolved the kernel adhesion and multi-crop incompatibility in the conventional imaging equipment of seed phenotyping. Dual-stage differential vibratory feeding was also integrated with sub-pixel edge extraction and adaptive morphological measurement. At the same time, high-throughput kernel counting and accurate two-dimensional morphological characterization were realized over the crop species with morphologically diverse kernels, without replacing mechanical components among crops. The hardware and software platform can provide the robust high-throughput phenotypic data support to evaluate crop germplasm resource, precision breeding selection, and functional genomic analysis of major cereal and legume crops.