SEM Pore Analysis

Automated pore size quantification from Scanning Electron Microscopy (SEM) images using adaptive Gaussian thresholding and contour analysis. Built for batch processing — recursively walks an input directory and analyses every image in a single run. In the original study, this pipeline was used to batch-analyse 62 SEM micrographs, detecting over 26,000 pores across multiple sample categories and time points, but the tool scales to any number of images.

Overview

In SEM micrographs of porous materials, pores appear as dark regions against a brighter solid matrix. This tool segments those dark regions, filters by size, and computes equivalent circular diameters for each detected pore, providing:

• Scalable batch analysis — point it at a directory tree with any number of images and get a consolidated CSV
• Per-image pore statistics (count, mean/median/SD diameter)
• Annotated images with green pore contours and red diameter labels
• Binary masks (white = pore, black = background)
• Publication-ready figures (composite overlay + summary bar chart)
• Reproducible CSV output with full analysis parameters

Algorithm

 SEM grayscale image (.tif)
        |
        v
 Crop metadata bar (bottom region)
        |
        v
 Gaussian blur (noise reduction)
        |
        v
 Adaptive Gaussian threshold (BINARY_INV)
   |-- Block size: neighbourhood for local threshold
   |-- C value:    constant subtracted from local mean
        |
        v
 Binary mask (dark pores -> white)
        |
        v
 Morphological opening (remove salt noise)
        |
        v
 Contour detection  ->  Filter by area & diameter
        |
        v
 Equivalent diameter:  D = 2 * sqrt(Area / pi)
        |
        v
 CSV statistics  +  Annotated images  +  Binary masks

The adaptive threshold computes a separate threshold for each pixel neighbourhood, making it robust to uneven illumination and contrast gradients common in SEM imaging.

Example Output

Composite Segmentation

Top row: original SEM micrographs. Bottom row: detected pore regions (magenta overlay).

Pore Diameter Chart

Annotated Image (with contours and diameter labels)

CSV Results

Image	N Pores	Mean (µm)	SD (µm)	Median (µm)
sample_1	455	3.37	4.81	1.71
sample_2	24	4.35	8.66	1.46
sample_3	341	3.77	4.36	2.29

Installation

git clone https://github.com/Harsh9005/sem-pore-analysis.git
cd sem-pore-analysis
pip install -r requirements.txt

Dependencies: NumPy, Pandas, OpenCV, Matplotlib

Usage

1. Pore Analysis (statistics + figures)

python sem_pore_analysis.py --input example_data/

2. Generate Annotated Images

python generate_annotated_images.py --input example_data/

Batch analysis

Both scripts recursively walk the input directory, so you can point them at a folder tree containing hundreds of images and every .tif file will be processed in one run:

# Analyse an entire dataset organised in sub-folders
python sem_pore_analysis.py --input /path/to/dataset/

# Only process 1000x magnification images
python sem_pore_analysis.py --input /path/to/dataset/ --magnification-filter 1000x

In the original study this pipeline batch-processed 62 images and detected >26,000 pores across multiple experimental conditions. The tool imposes no upper limit on dataset size.

Custom parameters

python sem_pore_analysis.py \
    --input path/to/images/ \
    --output path/to/results/ \
    --block-size 499 \
    --c-value 15 \
    --hfw 207.0 \
    --min-area 1.0 \
    --max-diameter 30 \
    --format png \
    --dpi 150

All options (sem_pore_analysis.py)

Argument	Default	Description
--input	required	Directory with .tif SEM images (recursive)
--output	<input>/analysis_output	Output directory
--block-size	899	Adaptive threshold neighbourhood size (px)
--c-value	10	Constant subtracted from local mean
--image-width	1536	Image width in pixels (for calibration)
--hfw	207.0	Horizontal field width in µm
--min-area	0.5	Minimum pore area in µm²
--max-diameter	50.0	Maximum pore equivalent diameter in µm
--blur-kernel	5	Gaussian blur kernel size
--crop-height	1000	Crop height to exclude metadata bar (px)
--morph-kernel	3	Morphological opening kernel size
--magnification-filter	none	Only process files containing this string
--dpi	300	Figure resolution
--format	png	Output format (png, tiff, pdf)

Output Files

analysis_output/
├── pore_analysis_results.csv      # Per-image pore statistics
├── composite_segmentation.png     # Side-by-side original + overlay
├── pore_diameter_chart.png        # Summary bar chart
├── analysis_parameters.txt        # Full parameter log
├── sample_1_mask.png              # Binary mask per image
├── sample_2_mask.png
└── sample_3_mask.png

annotated_output/
├── Annotated_sample_1.png         # Green contours + red labels
├── Annotated_sample_2.png
├── Annotated_sample_3.png
├── Mask_sample_1.png              # Full-size binary masks
├── Mask_sample_2.png
└── Mask_sample_3.png

How It Works

1. Metadata bar cropping — SEM software typically adds an information bar at the bottom of the image. The crop height parameter removes this region before analysis.

2. Gaussian smoothing — A small blur kernel reduces pixel noise without significantly affecting pore boundaries.

3. Adaptive Gaussian thresholding — Unlike global thresholding (e.g. Otsu's method), adaptive thresholding computes a local threshold for each pixel based on the weighted mean of its neighbourhood. This handles uneven brightness gradients across the field of view. The threshold is inverted so that dark pores become white in the binary mask.

4. Morphological opening — An erosion followed by dilation with a small kernel removes isolated bright pixels (false positives) while preserving the shape of genuine pore regions.

5. Contour analysis — OpenCV contour detection identifies connected white regions. Each contour's area is converted from pixels to µm² using the known calibration (image width in pixels / horizontal field width in µm). Contours outside the area and diameter bounds are discarded.

6. Equivalent diameter — For each valid pore, the equivalent circular diameter is computed as D = 2 * sqrt(Area / pi), representing the diameter of a circle with the same area as the detected pore.

Image Requirements

• Format: .tif (grayscale SEM micrographs)
• Calibration: Known image width and horizontal field width at the imaging magnification
• Default calibration: 1536 px / 207 µm (1000x magnification) — adjust via --image-width and --hfw for other setups

Author

Harshvardhan Modh

License

This project is licensed under the MIT License — see LICENSE for details.