Image classification and result analysis using JMP

[#1]

Hello JMP Community,

I have couple hundred images that were classified by different operators into three different groups (bad, good, and not sure). I would like to know what type of statistical analysis i should do (in JMP) on this data in order to implement the result into an image analysis program?

appreciate all your help,

[#2]

Hello @AFaiz,

Dpending on the topic, complexity of the task and details in the image that enables to classify the images, JMP may or may not be an appropriate tool.

One first step to get started on this topic and see if you need a more advanced approach could be to try the Image Analyzer Add-In. This Add-In enables to load images, and it will create data tables with RGB and HLS values for each pixels of your image. In order to process several images, scripting may be useful in order to automatize this process and concatenates the data of each image.

Using the distributions of RGB and/or HLS may help you to compare the images between different groups, and possibly detect a rule to automatically classify new images, thanks to a median/mean/std dev threshold, specific values of skewness or kurtosis, proportion zero values, range ...

I used this approach once for aluminum staining evaluation, as I wanted to know if it was possible to have a correlation between cotation scores (obtained from operators for the tesing of staining properties on aluminum from different formulations) and a data analysis approach thanks to images. I was very happily surprised to see that the data approach was able to match very closely the cotation scores obtained by experienced operators, and enable to have non-operator dependent scores.

However, this simple approach needs some prior fixed conditions for reproducibility and accuracy : fixed and defined background (white, grey or black depending on your objects), fixed position from the object, fixed lighting conditions, fixed camera settings, ...

If this approach is too simple for your task (different images size, complexity of objects in the images (use of different colors and sizes of the objects), different images quality/background, ...) a deep learning approach may be recommended. You might use the Python integration possibilities of JMP to use several dedicated libraries for images : OpenCV, Scikit-Image, ...

I hope this first answer will help you,