Research On Recognition Method Of Space Target Based On Image Features

With the further prosperous of the global development of space resources and the deep knowledge of the exploration of outer space, the importance of space target recognition has been gradually caused more and more researchers' attention. However, this area is still in early stages of development, there are many unresolved issues, so is necessary to perform extensive and in-depth research in the theory.This paper has discussed the following three issues: the existing methods in image preprocessing, feature extraction and target recognition. First, the functions and development of space target recognition as well as the basic methods of image preprocessing are introduced. According to imaging characteristics of visual light image and the complex environment of outer space, methods to smooth, enhance and restore image have been given. Then considering space target images may appear some features such as edges blurring, noises disturbance and deformation, a combination of gray, shape, and moment information extraction methods has been proposed for targets recognition. Fuzzy clustering algorithms have been investigated emphatically in this thesis. After introduction to the classical FCM algorithm, the FCM algorithm's limitation to data structure, which can only be applied to spherical or ellipsoidal clustering, has been pointed out. Followed by the description of the FCM algorithm, an improved algorithm, Kernel Fuzzy c-means clustering (KFCM), has been introduced. The algorithm utilized the idea of kernel methods in fuzzy clustering. In theory, The KFCM algorithm is able to solve the FCM algorithm's limitation. But in the actual, the KFCM algorithm will cause another problem that is easy to fall into local minimum. In this paper, through analyzing the algorithm's membership function, the reason of KFCM'S plunging into local minimum has been found out, and an algorithm of kernel clustering KFCM based on Voronoi space which is aiming at the reason of the question is brought out. The new algorithm can keep the systemic fuzzy, at the same time the crispness of the membership function is improved. On one hand, the objective function of mending can make the algorithm of KVFCM gain a more crispness membership, on the other hand, because Voronoi distance is used, the convergence and the robust of KVFCM algorithm are also improved. The experiments on real-world datasets which are generated from the hardware-in-the-loop simulation test platform of space targets recognition show that our proposed algorithms are more accuracy when compared to the FCM and the KFCM algorithm.