Automatic Clustering of Natural Scene Using Color Spatial Envelope Feature

A video scene can be defined as a fixed subdivision of a video, or a group of video frames having the same semantic contents. This paper presents a method to perform scene classification under unsupervised clustering environment. A holistic representation of the Spatial Envelope has been proposed to model the scene. One drawback of Spatial Envelope features is that it uses R, G, and B channels separately to extract features for processing. However, individual R, G, and B channels cannot describe color visual information of the image accurately. In this paper, a novel different color channel generated with Fibonacci lattice color quantization indexes is applied to generate Spatial Envelope features to address this drawback. An unsupervised clustering method named as Hyperclique Pattern-KMEANS (HP-KMEANS) is proposed to automatically select constraints for image clustering. Evaluation of the proposed feature extraction algorithm shows promising results for natural scene classification.