An Improved Fast ICA Algorithm for IR Objects Recognition

In this paper, an improved algorithm based on fast ICA and optimum selection for IR objects recognition is proposed. Directed against the problem that the Newton iteration is rather sensitive to the selection of initial value, this paper presents a one dimension search to improve its optimum learning algorithm in order to make the convergence of the results independent of the selection of the initial value. Meanwhile, we design a novel rule for the distance function to retain the features of the independent component having major contribution to object recognition. It overcomes the problem of declining of recognition rate and robustness associated with the increasing of training image samples. Compared with traditional methods the proposed algorithm can reach a higher recognition rate with fewer IR objects features and is more robust in different kinds of classes.