Extraction of image semantic features with spatial-range mean shift clustering algorithm

In recent years, the Bag-of-visual Words image representation has led to many significant results in visual object recognition and categorization. However, experiments show that the unsupervised clustering of primitive visual features tends to result in the limited discriminative ability of the visual codebook, since it does not take the spatial relationship between visual primitives into consideration. This paper aims at generating descriptive higher-order semantic features, which are extracted from visual word sets clustered by spatial-range mean shift and are a better representation for images. This method first uses mean shift algorithm to cluster visual words for an image from the spatial and color space, then uses FP-growth algorithm to mine the meaningful spatially concurrent groups of visual words in all images and regards the high frequency visual word combinations which can represent parts of objects as semantic features. The experiments on Caltech 101 dataset demonstrate that the proposed higher-order semantic features can achieve good results.