Co-occurrence Bag of Words for Object Recognition
Naeem Bhatti, Allan Hanbury · 2010
In the original 'bag of words' model image fea- tures are seen as independent and orderless. In the real world, relative co-occurrence of image features defines the spatial order in which they appear. This means the spatial relationship between image features has potential informa- tion to be exploited in a classification task. Our model attempts to discover the usefulness of the co-occurrence in- formation of image features with its extent ranging from local to global level targeting the image classification task. To establish the pair-wise relations between the features, we use the neighborhood of each image feature. To de- fine the extent of neighborhood of an image feature at a local level and to achieve scale invariance, the model uses a circular area based on the elliptical scale at which each image feature is detected. To range from local to global level, the model explores the extent of the neighborhood spatially at integral multiples of the circular area radius. In turn, each image is described by a histogram of pair-wise visual words and a concatenated histogram of independent visual words with pair-wise visual words. We develop a stop word removal technique to eliminate noisy redundant visual words from the collection. Experimental results show that the pair-wise co-occurrences of visual words alone and their augmentation with independent occurrences leads to a positive improvement in the classification accuracy.