Evolutionary Feature and Instance Selection for Visual Recognition
The aim of image mining is to automatically extract semantically meaningful information (or knowledge) from image data or to identify other patterns that are not explicitly stored in the images. In particular, image mining must be approached by image preprocessing. The preprocessing step in image mining deals with some conventional image processing tasks, such as segmentation and low-level feature extraction, where the mining result is heavily dependent on the preprocessing result, e.g. the identification of spatial objects and their relationships. From the KDD point of view, the processed image data that are represented by some feature vectors are used for the data analysis or mining stage. However, data preprocessing of KDD is usually based on feature selection and/or instance selection, in which the former focuses on dimensionality reduction and the later for data sample reduction. In this paper, we examine the effects of performing feature and instance selection by the genetic algorithm (GA) on image classification as a very important problem of image mining. The results based on two image datasets show that image mining can take advantage of feature and instance selection. Specifically, GA largely reduces the training dataset size by over 50% dimensionality reduction and near 50% training data sample reduction. In addition, the classifier trained by the reduced dataset can provide comparable accuracy with the baseline classifier without data preprocessing.
Keywords - Feature Selection, Genetic Algorithms, Image Mining, Instance Selection, Visual Recognition.