Unsupervised Bitstream Based Segmentation of Images

We consider here image segmentation as a problem of clustering texture features by frequency content. Specifically, we develop a low complexity algorithm for image segmentation that operates directly on the bitstream of JPEG compressed images. Using morphological filtering and watersheds, the algorithm effectively segments an image by combining areas of similar frequency content. Its low complexity and the fact that it does not require decoding makes it well-suited for distributed wireless sensor networks and image database search applications.