Human-inspired features for natural scene classification

Abstract

Scene classification has been the target of much research. Most psychological studies have agreed that humans perceive a scene first globally recognizing its category and then they localize and recognize objects. In previous work the same feature set were used in classifying both natural scenes and manmade scenes simultaneously. We suggest the use of different features for each. In this paper the proposed features for natural scenes classification are presented. The new proposed features are inspired from the way humans perceive and recognize scenes at a glance. Outdoor scenes global features such as openness, roughness, and dominant directions have been investigated and translated into a new feature set, focusing on characteristics that efficiently differentiate between natural scene sub-classes. The effectiveness of the proposed features is tested using two datasets consists of 4 natural scenes (coast, mountain, forest, and open country) and 6 natural scenes (the previous 4 scenes plus desert and waterfall scenes), the first dataset is a benchmark data set used for testing scene classification techniques. Results showed that a classification accuracy of up to 95% could be achieved using the proposed feature set.