Neural network based feature point detection for image morphing

In recent years, information security becomes more necessary than it used to be. Steganography is one of the effective mechanisms to protect ones privacy and secrets. Recently, we proposed a morphing based steganography, in which a morphed image is used as the cover image, the encryption key, as well as the stego-key. A key point to ensure the security is the “naturalness” of the morphed images. To obtain natural images through morphing, we may manually specify the feature points in the given reference images. This, however, is a very tedious task in practice. To increase the efficiency, we study in this paper automatic feature point detection based on neural networks (NNs). In this method, the difference between a sub-image A and a sub-image B is used as the input of the NN, and the output is the estimated difference between the coordinates of the centers of A and B. Thus, if B is centered by one of the feature points, and the NN is properly designed, we can obtain an estimated coordinate of the corresponding feature point directly from the NN output, given A-B as the input. This paper introduces the process for obtaining the training data and the teacher signals, and provides some initial experimental results to verify the proposed approach.