Evolution Constructed Features / Genetic algorithm for image classification

This repo provides an image classification algorithm based on a genetic algorithm for feature optimization and a RandomForestClassifiers-based ensempble classifier (adaboost) for prediction.

The initial genetic algorithm for image classification has been described in Lillywhite, et al.: http://dx.doi.org/10.1016/j.patcog.2013.06.002. The repo closely follows the proposed algorithm with some additional image processing kernels.

Installation

The repo is managed with poetry. To install the environment run

    poetry install

from the root directory.

Usage

To use the algorithm a library of images is required as well as a csv-file containing the labels and paths for each image. See classification.csv for more information.

There are two ways of training the classification model.

• Interactive: Load main.py as module and create an instance of the EvoFeatures-class. Call load_data and provide a path to the directory where the csv-file is located. Afterwards call fit and provide and output-directory.
• Script: Modify the if __name__-section at the end of main.py and modify line number 875 to point to the directory with the classification-csv-file. Then run the entire script.

In both cases the training ends with the creation of an adaboost.pkl file in the output directory.

Inference

• Load main.py as module and create an instance of the EnsembleClassifier-class, providing a path to the previously created adaboost-file.
• Load a set of images as a list of numpy arrays.
• Call the predict-method of the EnsembleClassifier-class and provide the list of images as argument