Finding the splitting vector for image resolution up-conversion

This paper proposes a new single-frame image up-conversion approach that uses prior information. The proposed method overcomes the drawbacks of the Kondo 2001 where the class membership of a local aperture depends only on the contents of itself, without taking any consideration of the contents of all the other training apertures, and all the features in an aperture share one single break point. We show that more effective classification can be achieved by making the break points adaptive to the content of all the training apertures. We propose to iteratively find the break points and the filtering coefficients in an expectation-maximization framework. The break points partition the entire training sample space into distinctive segments through axis-parallel splitting on each of the feature variables. The optimal coefficients for each segment are then obtained by LMS optimization. The partition-regression leads to a classification and regression tree. The proposed method exhibits improved performance over previous up-conversion approaches.