Machine-Vision Surface Porosity Analysis

Sat, 01 Nov 2025 00:00:00 +0000

Porosity affects the strength and durability of cementitious materials, and measuring it normally means cutting the specimen. I wanted to estimate it from images instead, so the same sample could still be used for other testing.

This started as a continuation of another student’s project. Trinh Vo had used MATLAB’s Image Processing Toolbox to estimate the porosity of thin plastic samples imaged by SEM. I extended that approach to cementitious materials, working with fly ash geopolymer concrete, and to images taken outside the SEM.

Specimens and imaging

The samples were fly ash geopolymer concrete, labeled by the compressive strength each cylinder reached at break. The images here were taken on a Hitachi S-4800 FESEM at 3 kV in secondary electron mode, at 700x with a 50 micrometer scale bar. I also ran the pipeline on ordinary macro photographs to see how it behaved outside the SEM.

Method

The approach is simple. Load the image, convert it to grayscale, threshold it into a binary image, and take the fraction of pore pixels.

img = imread(FILE_NAME);
gray_img = mean(img, 3);
binary_img = gray_img > THRESHOLD;
porosity = 1 - mean(binary_img(:));

Pipeline output for one specimen. Original micrograph, grayscale conversion, and the binary image the porosity value is calculated from.

Threshold sensitivity

The original code used a fixed threshold of 70. That value does not carry from one image set to the next.

On one SEM image, a threshold of 70 gave a porosity of 0.0832, and raising it to 75 gave 0.0937. On a macro photograph, a threshold of 70 returned 0.0000, and the value did not begin to move until roughly 72.

Initial work assumed that the choice of threshold does not significantly affect the result when the goal is ranking samples against each other. That holds only within a single imaging mode. The absolute values move with the threshold, and a macro image can return zero porosity purely as an artifact of it. Any use of this method needs a stated protocol for how the threshold was set.

What I changed

To reduce the dependence on a hand-picked number, I replaced the fixed threshold with imbinarize, which selects a threshold from the image histogram rather than from a constant in the script.

gray_img = im2gray(img);
BI = imbinarize(gray_img);
porosity = mean(BI(:));

I also tested adaptive thresholding for images with uneven lighting, and set up a script to try classifying pore regions with the Neural Network Toolbox, though I did not take that past initial setup.

Extension to object counting

The same front end was later reused for counting instead of area measurement. I developed this for oyster habitat rehabilitation research I worked on at Louisiana Tech, where the task was counting oysters in photographs of shell substrate.

Porosity only needs the fraction of dark pixels, so the pixels do not have to be grouped into anything. Counting does. Two steps were added after binarization. bwareaopen discards regions below a set pixel area, which clears the speckle that would otherwise be counted as objects. bwconncomp then labels the connected regions and returns how many there are. A hole-filled version is generated alongside them to check how the binary image is holding together.

BI = imbinarize(im2gray(img));
BW = 1 - BI; % objects dark, background white
denoise = bwareaopen(BW, 50); % drop regions under 50 pixels

cc = bwconncomp(BW);
cc2 = bwconncomp(denoise);
num = cc.NumObjects;
num_denoise = cc2.NumObjects;

Reporting the raw count alongside the denoised count shows how much of the result came from noise.

Counting output for a substrate photograph, showing the original image and the binary, denoised, and hole-filled stages. This run returned a count of 337.

Dense substrate is the hard case. Shells that touch merge into a single connected region and undercount, which is visible in the binary panels above as large joined masses rather than separate objects. The count is most reliable when the objects are well separated.

Limitations

This measures surface porosity in the imaged field, not bulk porosity. It cannot see closed pores below the surface, and it is sensitive to surface preparation, lighting, and magnification. I treat it as a fast comparative measure between specimens imaged the same way, not as a replacement for a standard porosity test.