Using sub-dictionaries for image representation based on the bag-of-visual-words approach

Bag-of-Visual-Words (BoVW) is a well known approach to represent images for visual recognition and retrieval tasks. This approach represents an image as a histogram of visual words and the dissimilarity between two images is measured by comparing those histograms. When performing comparisons involving a specific type of images, some visual words can be more informative and discriminative than others. To take advantage of this fact, assigning appropriate weights can improve the performance of image retrieval. In this paper, we developed a novel modeling approach based on sub-dictionaries. We extracted a sub-dictionary as a subset of visual words that best represents a specific image class. To measure the dissimilarity distance between images, we take into account the distance of the histogram obtained using the visual dictionary and the distances of the sub-histograms obtained by each sub-dictionary. The proposed approach was evaluated by classifying a standard biomedical image dataset into categories defined by image modality and body part and also natural image scenes. The experimental results demonstrate the gain obtained of the proposed weighting approach when compared to the traditional weighting approach based on TF-IDF (Term Frequency-Inverse Document Frequency). Our proposed approach has shown promising results to boost the classification accuracy as well as the retrieval precision. Moreover, it does that without increasing the feature vector dimensionality.