Research On Underwater Target Recognition Based On BP Neural Network

In recent years, with the development of ocean exploitation and the need of national defenses. Autonomous Underwater Vehicle (AUV) has been widely applied. Complexities and uncertainties of working environments make the vision system of AUV’s stand out especially. This paper carries out the research on underwater optical vision system with a subject named "The motion control of Underwater Vehicle".Image segmentation and pattern recognition technologies for underwater objects are two important links in underwater optical vision system. The purpose of this thesis is to investigate a set of image segmentation and target recognition algorithms with good real-time performance and robustness through subject interrelated research, and then to construct an optical vision based underwater target recognition system. The work of this paper is mainly concentrated on the image enhancement, image segmentation, feature extraction and recognition of the underwater images.First, as for disadvantages of the underwater images arising from the low contrast and the disequilibrium of histogram distribution, a image enhancement algorithm based on maximal entropy theory is proposed, this algorithm enhances the contrast of the image, and the target and background are much more evident.Second, due to the effect of the uneven illumination, OSTU improved by GA is proposed, this algorithm overcomes effectively the defect of the uneven illumination, and improves greatly the speed and get the better quality at image segmentation.Third, because of the influence of scale factor on Hu’s moment invariant features in discrete, six new moment invariants with both scale, translation and rotating invariance are constructed, these new moment invariants have better clustering effect.Finally, to deal with the deficiency of the BP neural network, such as the low speed of convergence, traps into local optima easily, the moment factor and the self-adaptive adjusting learn rate are introduced, all pixels and the new moment invariants are taken as input data for BP individually, the experimental results show that the improved BP neural network taken the new moment invariants as input, gives a high accuracy for underwater object recognition with enough robustness.In conclusion, underwater optical vision system is build through the study on the image enhancement, image segmentation, feature extraction and recognition of the underwater images. It is shown that the algorithms provided in the thesis are effective and instructive to go on the in-depth study of underwater optical vision technology.