Substantiating the YOLO11 architecture for determining the fractional composition of winter wheat grain mixtures

This study’s object is the process of determining the fractional composition of winter wheat grain mixtures using computer vision and deep learning methods. The basic task that needs to be solved is the high complexity, subjectivity, and speed of determining the fractional composition of grain by conventional methods.

The results obtained demonstrate the successful training and comparative analysis of several YOLO11-seg instance segmentation models on a specialized dataset deployed on the NVIDIA Jetson Orin platform. In particular, the YOLO11m-seg model with an image size of 640 × 640 pixels achieved the optimal compromise between accuracy and speed, achieving a Mask mAP50-95 index of 0.558 at an output speed of 62.5 ms/image. Training the YOLO11n-seg 1280 × 1280 model provided the best average segmentation accuracy (Mask mAP50-95 0.640) by increasing the performance of identifying objects of "complex" classes, which are key for accurate determination of the fractional composition.

The results have made it possible to solve the problem under consideration through the empirically justified choice of architecture. Unlike hypothetical approaches, the study provides specific quantitative data on the performance of different YOLO11-seg architectures. That allowed for a reasonable selection of the model that best meets the requirements for accuracy and speed for practical deployment, solving the problem of uncertainty in the choice of architecture.

The findings create the basis for automating grain quality control, increasing its efficiency, objectivity, and also significantly reducing labor intensity by automating routine operations. For practical use of the system, it is necessary to ensure stable lighting conditions, as well as the presence of a digital camera, a computer, and appropriate software