Granulometry Monitoring System

Real-time industrial system for detecting, counting, and measuring granules (~5 mm) on a production conveyor. YOLO-based detection pipeline for real-time inference Estimated real-world object size using camera calibration and projection geometry Integrated GigE Vision industrial camera system Exported model to ONNX for optimized inference performance Implemented real-time alert system (visual + audio signals) Logged measurements into structured reporting system

Fusarium Detection in Wheat (Spectral Analysis)

Computer vision system for detecting Fusarium infection in wheat using spectral image analysis. Developed RGB-to-HSV transformation for spectral feature extraction Built mathematical model mapping HSV Hue to wavelength (380–740 nm) Analyzed spectral differences between healthy and infected plants Designed wavelength-based disease detection approach Awarded research grant by Novgorod Innovation Development Center

Microchip Defect Detection

Automated optical inspection system for detecting microchip defects in high-resolution (2K) microscope images. Designed a two-stage YOLOv11 pipeline (ROI extraction + defect detection) Implemented OpenCV preprocessing for noise reduction and normalization Reduced computational load via ROI-based inference strategy Annotated and augmented dataset (1,500 → 2,500 images) using CVAT Exported results to structured CSV reports for industrial analysis Awarded the “Student Startup” grant

Boiling Layer Level Monitoring System

Real-time computer vision system for monitoring boiling layer level in industrial tanks. OpenCV-based pipeline for boundary detection in noisy environments Extracted upper boundary position in real time from video stream Converted pixel coordinates to real-world measurements using calibration Implemented threshold-based anomaly detection system Integrated into unified multi-system industrial monitoring interface

Handwritten Manuscript Attribution & Dating

Tool for attributing, localizing, and dating 18th-century handwritten manuscripts from Russian theological school archives, aimed at supporting historians in identifying undocumented materials. Designed a YOLOv11 pipeline to segment manuscripts into word-level fragments Used CLIP (ViT-B-32) embeddings to represent each fragment and compare it against a database of manuscripts with known provenance via similarity search Returned similarity scores to estimate likely origin, time period, and author of unidentified manuscripts Comparable in spirit to Russia's "Digital Peter" handwriting recognition project, using a retrieval-based approach rather than direct transcription Project was not brought to completion, as the lab conducting the research was shut down