Evolution of Statistical Descriptors for the Image Recognition of Natural Sceneries by Means of Genetic Programming for CBIR Improvement
Juan Villegas-Cortéz, Carlos Avilés‐Cruz, Arturo Zúñiga‐López, Salomón Cordero-Sánchez, Francisco Fernández de Vega, Francisco P. Chávez · 2018
The rise of the Internet involves the simultaneous growth of the number of images in it. This amount of images comprises roughly more than half of the Internet content. This situation poses an open problem: how to recognize images from their own analysis without the use of labels describing their content or the analysis of documents or pages where the images being analyzed are appearing. In this work we present a novel approach for improving both the analysis technique and the classification of images of natural sceneries by using the content based image retrieval (CBIR) methodology which is applied for visual search. This improvement consists of the multigene evolution by means of genetic programming of new statistical texture descriptors in accordance with the type of scenery under analysis, the amount of the descriptors being used and the number of images. The percentage of recognition reaches up to 85% for natural scenery images considering 5 classes, showing a satisfactory improvement with new evolved solutions.