Synthesizing feature agents using evolutionary computation

Abstract

In this paper, genetic programming (GP) with smart crossover and smart mutation is proposed to discover integrated feature agents that are evolved from combinations of primitive image processing operations to extract regions-of-interest (ROIs) in remotely sensed images. The motivation for using genetic programming is to overcome the limitations of human experts, since GP attempts many unconventional ways of combination, in some cases, these unconventional combinations yield exceptionally good results. Smart crossover and smart mutation identify and keep the effective components of integrated operators called “agents” and significantly improve the efficiency of GP. Our experimental results show that compared to normal GP, our GP algorithm with smart crossover and smart mutation can find good agents more quickly during training to effectively extract the regions-of-interest and the learned agents can be applied to extract ROIs in other similar images.

Introduction

Image segmentation and labeling of terrain regions is an important task in remote sensing application (Mvogo et al., 2000; Jeon et al., 2002; Katartzis et al., 2001; Segl and Kaufmann, 2001; Acqua and Gamba, 2001; Dong et al., 2001). The quality of this task is primarily dependent on the data and features used as input. There are many kinds of features that can be used, and the question is what are the appropriate features or how to synthesize features, particularly useful for segmentation/labeling, from the primitive features extracted from images. Usually, there are almost infinite ways of combining primitive features to form synthesized composite ones. It is the human image experts who, relying on their rich experience, figure out a smart way of combination. The task of finding a good combination is equivalent to finding a good point in the search space of agents formed by the combination of primitive operations (also called primitive operators) on images.

However, limited by their speed, previous experience or even bias, the human experts can only try a very limited number of conventional combinations. Genetic programming (GP), on the other hand, may try many unconventional combinations that may yield exceptionally good results. Also, genetic programming can explore a much larger portion of agent space due to the inherent parallelism of GP and the speed of computer. The search performed by GP is guided by the goodness of agents in the population. As the search proceeds, GP will gradually shift the population to the portion of the space that contains good agents. The crossover operation is the major mechanism used by GP to search the agent space and the mutation operation is employed to increase the diversity of the population to avoid the premature convergence. However, in the traditional crossover and mutation, their locations are randomly selected, leading to disrupting the effective components (subtree in this paper) within agents and thus greatly reducing the efficiency of GP. It is very important for GP to identify and keep those effective components.

In this paper, we use genetic programming to generate feature agents for terrain labeling. The individuals in our GP-based learning are agents represented by binary trees whose internal nodes represent the pre-specified primitive operators and the leaf nodes represent the original image or the primitive feature images generated from the original image. The primitive feature images are pre-determined, and they are not the output of the pre-specified primitive operators. After applying an agent on the original image or the primitive feature images, the output image of the agent is segmented to yield a binary image or mask. The binary mask is used to extract the particular kind of terrain regions from the original image. To improve the efficiency of GP, we propose smart crossover and smart mutation to identify and keep the effective components of an agent.