Abstract:

Objective To address the limitations of traditional methods for detecting starch granule size in mature maize kernels (such as laser diffraction particle size analysis), including cumbersome sample preparation, low efficiency, and poor accuracy, this study aimed to develop an efficient and accurate automatic analysis tool. This tool is intended to meet the demand for starch granule size measurement in maize yield and quality improvement, as well as in the production of specialty maize starch. Method Based on the YOLOv10 object detection algorithm, an automatic detection model for maize kernel starch granules was established and trained. A granule size analysis system was developed using the Django framework. Mature kernels from six representative maize inbred lines (B73, B104, KN5585, MO17, RP125, W22) were selected. Scanning electron microscopy (SEM) images of their endosperms were acquired. Starch granule size of the six inbred lines endosperms was analyzed using the developed system. Result Model evaluation indicates that the constructed starch granule detection model has an accuracy rate of 0.812, a recall rate of 0.843, and an average precision (AP) of 0.895. The developed system supports batch uploading of kernel endosperm SEM images by users. It utilizes the trained model to automatically identify starch granules within the images, analyzes and calculates the size data for each detected granule, and ultimately outputs the results in an Excel file. Analysis of starch granules from the six maize inbred lines revealed significant differences in mean granule diameter among them. MO17 showed the largest mean starch granule diameter (15.46 µm), while RP125 presented the smallest (10.61 µm). Compared with manual starch granule size measurements from SEM images, the system-generated granule size data demonstrates reliability, offering significant advantages in labor-saving and time-efficiency, with a substantial improvement in work productivity. Conclusion This study developed a maize kernel starch granule size analysis system based on the YOLOv10 object detection algorithm and the Django framework. The system enables the rapid, batch, and accurate automatic identification of starch granules in maize endosperm SEM images and analysis of their size. It effectively overcomes the limitations of traditional methods, providing a convenient, efficient, and accurate granule size analysis tool.

Key words: maize, endosperm of kernel, starch granule recognition, starch granule size detection, YOLOv10, deep learning, image recognition, Django framework