Development and testing of an online model for corn seed quality analysis using deep learning

The study was aimed to develop and test an online model based on YOLOv8n for detecting mechanical and biological damage to corn seeds based on RGB images, adapted for mobile applications. The work was carried out in 2025 in laboratory (artificial lighting with 4000 K fluorescent lamps) and field (natural daylight) conditions of the Cis-Ural forest-steppe zone of the Republic of Bashkortostan. The object was seeds of the Ross 190 MV and Ross 199 MV corn hybrids; shooting was taken with the main camera of a smartphone (64 MP, f / 1.7), 600 JPEG images with a resolution of 3000 × 4000 pixels were obtained. The dataset is annotated into two classes (healthy and damaged seeds) in Roboflow and divided into training (70 %), validation (15 %) and test (15 %) sets; the following augmentation was used: rotation of ±30°, change of brightness and contrast, scaling, reflection. The YOLOv8n model was trained for 150 epochs (batch 16, AdamW optimiser, initial learning rate of 0.001) on the Roboflow cloud platform. On the test set (90 images), the mAP@0.5 indicator was equal to 0.95; the detection accuracy of healthy seeds was 97.0 %, damaged ones – 93.0 %; recall for damaged ones was 87.5 %, accuracy – 85.4 %. The inference time via Roboflow Inference API was 150-300 ms with stable Wi-Fi (>50 Mbps) and up to 1-2 s with unstable cellular communication (3G/4G). The developed model ensures high accuracy and acceptable analysis speed, and is suitable for express quality control of seeds at storage and reception sites. Limitations include errors due to suboptimal lighting and seed overlap. The approach is scalable to other crops without changing the architecture.