Techniques for efficient and effective transformed image identification

In many applications, one common problem is to identify images which may have undergone unknown transformations. We define this problem as transformed image identification (TII), where the goal is to identify geometrically transformed and signal processed images for a given test image. The TII consists of three main stages – feature detection, feature representation, and feature matching. The TII approach by Lowe [D.G. Lowe, Distinctive image features from scale-invariant keypoints, Int. J. Comput. Vision 60 (2) (2004) 91–110] is one of the most promising techniques. However, both of its feature detection and matching stages are expensive, because a large number of feature-points are detected in the image scale-space and each feature-point is described using a high dimensional vector. In this paper, we explore the use of different techniques in each of the three TII stages and propose a number of promising TII approaches by combining different techniques of the three stages. Our experimental results reveal that the proposed approaches not only improve the computational efficiency and decrease the storage requirement significantly, but also increase the transformed image identification accuracy (robustness).