OptimalColloidNet: A Deep Neural Network-Based Framework for Colloidal Particle Detection in Microscopic Images

OptimalColloidNet is a deep learning framework released by aryanarora06 on GitHub (release date: 2026-05-30, link: https://github.com/aryanarora06/OptimalColloidNet), designed to address the robustness and accuracy issues of colloidal particle detection in microscopic images. This framework integrates attention mechanisms, multi-scale feature extraction, and sub-pixel center estimation techniques to achieve accurate particle detection under various conditions. It includes a training module (optimalcolloidpython.ipynb) and an inference module (analyze_video.py), supporting synthetic data training and real video analysis.

In materials science, chemistry, and biomedical research, microscopic imaging analysis of colloidal particles is crucial. Traditional methods rely on handcrafted features and threshold algorithms, whose accuracy drops significantly under complex conditions such as uneven particle density, lighting changes, or particle overlap. While existing deep learning solutions can automatically learn features, they lack scale adaptability and struggle to provide sub-pixel-level localization, limiting applications in precise measurements (e.g., particle trajectories, size distribution).

OptimalColloidNet's core technologies include:

1. Attention Mechanism: Focuses on potential particle regions to improve detection efficiency and accuracy;
2. Multi-scale Feature Extraction: Captures small-scale details (edges) and large-scale context (neighborhoods) to adapt to particles of different sizes;
3. Sub-pixel Center Estimation: Infers precise positions within the pixel grid to provide accurate data for subsequent analysis;
4. Test-Time Augmentation (TTA): Aggregates prediction results from 8-fold transformations to reduce uncertainty and improve stability.

The training process adopts a synthetic data generation strategy, dynamically generating 2000 images that simulate real experimental conditions (varying particle density, background noise, uneven lighting). Training configuration: early stopping mechanism (patience=8), maximum 60 training epochs, taking 20-40 minutes on a T4 GPU. After training, the best model weights are saved to colloid_output/best_checkpoint.pt, and training/validation loss curves are generated.

This framework has practical scientific research value and provides Kaggle Notebook support (cloud GPU acceleration, low technical threshold). Application scenarios include:

• Colloidal dynamics research: Track particle movement trajectories and analyze diffusion and interactions;
• Material characterization: Automatically count particle size distribution and evaluate sample uniformity;
• Biomedical imaging: Detect positions and morphologies of biological samples such as cells and microspheres;
• Industrial quality control: Automated detection of colloidal products.

Usage recommendations and tuning:

• Insufficient VRAM: Adjust batch size from default 8 to 4;
• Detection Threshold: If no particles are detected, reduce DETECT_THRESHOLD (from default 0.15 → try 0.10);
• Scale Adaptation: If Hough circle detection fails, manually set scale_factor.

OptimalColloidNet is a significant advancement in the field of particle detection in microscopic images. Integrating attention mechanisms, multi-scale feature extraction, and sub-pixel localization, it combines high accuracy and generalization ability, providing a practical open-source tool for colloidal science, materials research, and biomedical imaging. Future expansion directions include introducing advanced backbone networks, supporting 3D volume data, and integrating into larger image analysis pipelines.