Scene Classification using Regional and Nearest Neighbors of Local features

This paper aims to model the spatial neighbors of local image features in regular and irregular contextual regions. By establishing the spatial neighborhoods of local image features, we incorporate them into the classical Bag of Words model by making an extension to it. To obtain more efficient representations of images based on spatial neighborhood information of local image features, we propose three neighborhood models: X-regional neighbors model, K-nearest neighbors model, and Fixed radius nearest neighbors model. For each of the proposed neighborhood models, we establish the spatial relationships between the local features in terms of the Euclidean distance in the specified neighborhood and capture the neighborhood information of the local features in a pair-wise fashion. Evaluation is obtained by performing scene classification experiments on the 15 scene categories dataset. The results show that the proposed models produce better results than the existing bag of words model and context models from literature in terms of classification accuracy.