Toward an automatic exploration of algorithm space to speed up image annotation for applications in scientific image understanding
Dirk Colbry · 2023
This project demonstrates a new image analysis approach to help researchers navigate the large search space of image annotation algorithms and their hyperparameters. A prototype of this approach (called SEE-Segment) demonstrates the exploration of image segmentation using a single training example, although it is expected to have overfitting and generalization problems. SEE-Segment implements a solution to the Combined Algorithm Selection and Hyperparameter (CASH) Optimization problem and uses a genetic algorithm to search an "algorithm space" for solutions that may work for the provided problem. The output generated by SEE-Segment is standalone code that is not dependent on SEE-Segment and can easily be copied and pasted into researchers’ existing data processing workflows. Thus, SEE-Segment can also be seen as an educational tool that keeps the researchers "in the loop" and helps them to understand and use the segmentation algorithms that are available to them. This work will present preliminary results of running SEE-Segment on a variety of scientific image datasets and outline procedures to extend the approach to other annotation workflows by building an Image Annotation Grammar.