Evolutionary multi-objective optimization of trace transform for invariant feature extraction


Trace transform is one representation of images that uses different functionals applied on the image function. When the functional is integral, it becomes identical to the well-known Radon transform, which is a useful tool in computed tomography medical imaging. The key question in Trace transform is to select the best combination of the Trace functionals to produce the optimal triple feature, which is a challenging task. In this paper, we adopt a multi-objective evolutionary algorithm adapted from the elitist non-dominated sorting genetic algorithm (NSGA-II), an evolutionary algorithm that has shown to be very efficient for multi-objective optimization, to select the best functionals as well as the optimal number of projections used in Trace transform to achieve invariant image identification. This is achieved by minimizing the within-class variance and maximizing the between-class variance. To enhance the computational efficiency, the Trace parameters are calculated offline and stored, which are then used to calculate the triple features in the evolutionary optimization. The proposed Evolutionary Trace Transform (ETT) is empirically evaluated on various images from fish database. It is shown that the proposed algorithm is very promising in that it is computationally efficient and considerably outperforms existing methods in literatures.