GOOFeD: Extracting Advanced Features for Image Classification via Improved Genetic Programming
Stanton R. Price, Derek T. Anderson, Steven R. Price · 2019
Feature extraction is widely considered one of the most critical components to classification performance in computer vision. In the past, human-designed features, such as the histogram of oriented gradients, were used for extracting statistically rich features. Recently, there has been a movement away from human-designed features to machine-learned features. Herein, we propose a novel genetic programming (GP) approach, coined GOOFeD, to automatically generate discriminative-rich features for image classification. This is achieved by greatly advancing GP in three ways: (1) promoting population diversity and redundancy removal, (2) introducing a unique adaptive mutation approach, and (3) controlling tree bloat through a new crossover technique. These improvements also lead to a population size required for learning that is smaller than that commonly used in the literature. To assess performance, GOOFeD is tested on the MIT urban and nature scene data set and a real-world buried explosive hazard data set. Experiments verify that, in terms of classification accuracy, GOOFeD outperforms many of the state-of-the-art human-designed features and feature learning techniques.