Using machine learning to find an optimal set of parameters for a given segmentation algorithm.
In the "classical" case of machine learning, in the training phase, the data set is constant and the model fit — produces a weight vector that maps the data set to the tag of each image label.
Now let's assume a given segmentation problem X, which is done using a given classic segmentation algorithm Y (classic, not Deep-Learning). The goal is to find an optimal parameter set for the Y algorithm under the set of ground truth segmentation. (Motivation: every segmentation algo. Have tuned, parameters, we want to learn them not to fine tune them)
I think about the two approaches of making the this:
- Offline – Extraction of General properties – let's say Haralick texture features and try to fit a model connecting between the parameters of the segmentation algorithm Y to the Haralick texture features.
- Online – Select random parameters for the Segmentation Y algorithm. Perform a segmentation for those parameters and a specific delta. Calculate the error and then updated the parameters accordingly.
Any example/reference(paper) for the "online" approach would be welcome.
Best Answer
You can make more intelligent search strategies for finding your ideal parameters. For example you can use the Sequential Model-Based Optimization algorithm (https://www.cs.ubc.ca/~hutter/papers/10-TR-SMAC.pdf), for identifying promising points in your hyperparameter space. That explores regions of the hyperparameter space, that are unexplored and promise good results (based on your previous experiments).