High-level image representation-based on Gestalt theory for image classification

Feature extraction and representation are crucial stages for image classification. However, the derived features might not be robust discriminators, which radically alter image classification results. In this paper, we present an enhanced Bag of visual Words representation for better image classification. We are inspired by the Gestalt laws of grouping.So, our design goals are:•Introduce the spatial information into Bag of Words representation by grouping the aligned visual words (high order features),•Preserve the spatial relationships among the visual words groups against different geometric transformations,•Achieve the high-level representation of visual words by extracting visual phrases. Theses phrases convey the semantic meaning of object parts.The proposed approach has been evaluated, in terms of classification accuracy, on Caltech101 image database. Experimental results indicate that the classification performances are improved by using visual phrases.