COMPUTER VISION AND AUTOMATED IMAGE CHARACTERIZATION FOR USE IN TEXTURE AND FRACTURE ANALYSIS

Texture and fracture analysis of geologic images often requires manual tracing or point counting. Programs such as the Adobe Suite, FracPac, and ImageJ can edit images and analyze traced features. However, these programs still require researchers to invest significant time in the identification and tracing of grains and fractures before analysis can begin. The goal of this project is to create a program capable of completing two tasks: (1) automatically differentiate images by composition and provide analysis of texture and (2) accurately identify and trace fractures as vector objects through the use of machine learning.

To map the image by composition, the program identifies distinct groupings by color in X-ray maps or backscatter images. This is done through the use of the K-means algorithm, which classifies pixels into a set number of groups identified by the user. These groupings of pixels can then be analyzed to measure percent composition, grain size, and grain area relative to perimeter using additional OpenCV functions.

We plan to map fractures in images using machine learning in python. The program will be trained to distinguish fractures with a training dataset of manually traced fractures in three overlain images of sample areas: backscatter, color-CL, and manual vector tracings. Comparison of the backscatter and color-CL images of healed fractures in microcline may allow the program to distinguish fractures identified in the training data and apply this capability to image sets input into the program. These fractures are often hard to distinguish from exsolution lamellae, providing a robust test of the program.

Initial tests of the texture-analysis feature were conducted using low-resolution SEM-EDS X-ray maps. The composition of the image by group calculated by the program is relatively inaccurate, with values ranging ±9% compared to manually traced maps. However, the low resolution of the test image results in image noise which significantly affects the accuracy of the program. When run on a high-resolution test image created using Adobe Photoshop, the program is accurate to ±1%, implying that the program will be far more precise with higher resolution X-ray maps. Further work is needed to solve issues surrounding the accuracy with lower resolution images.