Analyzing images of thin sections to extract knowledge is a complex task made possible today by the emergence of deep learning for computer vision.
Our objective is to automate the recognition and extraction of small objects, such as microfossils, in thin sections in order to facilitate and accelerate the interpretation of samples by geologists.
We use algorithms based on convolutional neural networks specially designed to limit the number of annotated images required for learning. Thus, we can quickly have a system capable of improving itself by detecting the elements missed during the annotation phases.
The resulting model is able to accurately segment new images to delineate the objects of interest for the geologist.
This method could quickly be applied to any type of image and knowledge, even if the annotation base is small. Besides, a complementary approach relies on the adaptation of object detection algorithms to rock images.