Rotation-invariant features based on steerable transforms with an application to distributed image classification

In this paper, we propose a new rotation-invariant image retrieval system based on steerable pyramids and the concept of angular alignment across scales. First, we define energy-based texture features which are steerable under rotation, i.e., such that features corresponding to the rotated version of an image can be easily obtained from the features of the original (non-rotated) image. We also propose an approach to measure similarity between images that is robust to rotation; images are compared after being aligned in angle. The retrieval process is performed by means of a decision tree classifier where the angular alignment is performed at each node in the tree. To demonstrate the effectiveness of our system we consider a distributed image classification system, where the feature encoder and the classifier are physically apart and thus features are compressed before being transmitted. Our results of retrieval performance versus rate show a clear gain with respect to a wavelet transform (as an example, for the same rate, the retrieval precision is increased from 40% to 65%).