Multi-class classification of objects in images using principal component analysis and genetic programming

This work presents a methodology for using Principal Component Analysis (PCA) and Genetic Programming (GP) for the classification of multi-class objects found in digital images. The image classification process is performed by using features extracted from images, through feature extraction algorithms, reduced by PCA and labeled by similarity comparing with other previously classified objects. GP uses two sets of elements: terminals, composed by the features extracted by PCA; and non-terminals, composed by algebraic operations. The fitness function was defined by the product of sensibility and specificity, two performance measures. A penalty term is also used to decrease the number of nodes of the tree, while minimally affecting the quality of solutions. The proposed approach was applied to set of 2739 digital images divided into objects representing airplanes, motorbikes, background from google, faces and watch classes, provided by the Caltech101 image database. The proposed approach was compared with SVM, Naïve Bayes and C4.5. Results suggest that the approach PCA+GP is able to evolve solutions for the problem as a simple classification rule with true positive rate above 70%. Additionally, we observe that PCA+PG obtained results slightly better than SVM and C4.5, besides these methods give a result that is not comprehensible by humans.