Created by W.Langdon from gp-bibliography.bib Revision:1.4910
First, this dissertation shows the efficacy of GP and CGP in synthesizing effective composite operators and composite features from domain-independent primitive image processing operations and primitive features for object detection and recognition. Based on GP and CGP's ability of synthesizing effective features from simple features not specifically designed for a particular kind of imagery, the cost of building object detection and recognition systems is lowered and the flexibility of the systems is increased. More importantly, it shows that a large amount of unconventional features are explored by GP and CGP and these unconventional features yield exceptionally good detection and recognition performances in some cases, overcoming the human experts' limitation of considering only a small number of conventional features.
Second, smart crossover, smart mutation and a new fitness function based on minimum description length (MDL) principle are designed to improve the efficiency of genetic programming. Smart crossover and smart mutation are designed to identify and keep the effective components of composite operators from being disrupted and a MDL-based fitness function is proposed to address the well-known code bloat problem of GP without imposing severe restriction on the GP search. Compared to normal GP, smart GP algorithm with smart crossover, smart mutation and a MDL-based fitness function finds effective composite operators more quickly and the composite operators learned by smart GP algorithm have smaller size, greatly reducing both the computational expense during testing and the possibility of overfitting during training.
Finally, a new MDL-based fitness function is proposed to improve the genetic algorithm's performance on feature selection for object detection and recognition. The MDL-based fitness function incorporates the number of features selected into the fitness evaluation process and prevents GA from selecting a large number of features to overfit the training data. The goal is to select a small set of features with good discrimination performances on both training and unseen testing data to reduce the possibility of overfitting the training data during training and the computational burden during testing.",
Supervisor Bir Bhanu
Senior Research Engineer, Trend Micro Inc UMI Number: 3096772 ProQuest Order No. 3096772 OCLC: 53984756",
Genetic Programming entries for Yingqiang Lin