A developer created a computer vision system to monitor buckwheat quality on production lines, specifically to analyze images from photo separators. The model identifies three categories of grains: good, bad, and grey, helping to optimize separator settings and reduce product loss.
The dataset consisted of 24 images captured by a linear camera, each measuring 2716 by 512 pixels and containing between 200 and 800 grains. Due to the small sample size, the developer generated 500 synthetic images for training, using YOLOv11 and YOLO26 small architectures.
Training and annotation were performed on the free platform picva.ru, which supports GPU-based training and model export. The final model was exported in ONNX format and embedded into a single 67.4 MB HTML file using base64 encoding, allowing it to run locally in a browser without internet or Python.
The solution utilizes onnxruntime-web 1.20.1 to perform inference via WebAssembly. This approach allows laboratory staff to process batches of photos offline, providing immediate statistics on grain quality distribution without requiring software installation or external server access.
