Abstract: In order to narrow the semantic gap between low-level visual features and high-level semantic concepts in image categorization, the clustering information from a density clustering algorithm and the characteristics of multi-instance learning framework in distinguishing ambiguous object are exploited. An image categorization method is proposed using Density Clustering on Region Features and Multi-Instance Learning (DCRF-MIL). First, the DCRF-MIL divides each image into number of regions and relines up all regions into a collection; then it uses a density clustering algorithm to learn the potential distribution information of the region features in the collection. Second, it treats an image as a bag and the regions as instances in the bag. Based on the cluster distribution information of region features, the bag is mapped into a vector in the cluster distribution space. Finally, a support vector machine classifier is constructed to predict the class label of the unlabeled image. The experiments on the Corel image data set and MUSK molecular activity prediction data set show that the DCRF-MIL method has high classification accuracy and it is easy to select its parameters.

Key words: artificial intelligence, image categorization, multi-instance learning, density clustering, support vector machine