UMP Dot Matrix Pattern Recognition with Object Detection

Micro dot detection · YOLOv8-style architecture · 0/1 Classification · Production-grade inference

Overview

This project focuses on detecting micro dot patterns on a unit surface — specifically circular dot clusters and a long bar pattern — and classifying each detected unit pattern as 0 or 1. Due to confidentiality, the exact product and pattern meaning are not disclosed; the scope of my work was robust detection and classification of the pattern units under real production conditions.

🔍 Object Detection ⚫ Small Object PyTorch OpenCV C#
View GitHub Code

Problem

  • Targets are very small objects (tiny dots / short segments), making detection challenging.
  • High risk of missed detections due to:
    • Low contrast / reflective surfaces
    • Sensor noise
    • Small object scale relative to full image
    • Dense patterns (many nearby objects)

Solution

  • Implemented an object detection pipeline based on a YOLOv8-style architecture, customized for small-object performance.
  • Key improvements included:
    • Higher input resolution / multi-scale strategy
    • Feature refinement for small targets (stronger shallow features, better fusion)
    • Anchor / detection head tuning for tiny objects
    • Augmentation strategy tailored to pattern consistency (blur, noise, brightness, rotation)
    • Optional tiling / ROI-based inference to preserve tiny details when needed

Output

  • Detects each unit pattern (dot region / bar region)
  • Assigns 0/1 label per detected unit
  • Produces overlay results for verification (bounding boxes + confidence)

Implementation

Language: Python, C#

Libraries:

PyTorch OpenCV NumPy

Framework: YOLOv8-style detection framework (customized)

Outputs: Bounding boxes · Confidence scores · 0/1 labels · Overlay visualization

Result

  • Reliable detection of dense, tiny pattern units with production-friendly inference speed.
  • Reduced manual inspection by providing consistent, measurable detection outputs.