Study On Vision-based Underwater Object Recognition

With the increasing consumption of land resources, the development of marine resources and underwater operation tasks become increasingly frequent. At the same time, underwater robot gets rapid growth, which cannot be inseparable from the target detection and identification. As a particular application field of target recognition, underwater target recognition consists of not only the general steps of object recognition, but also the image processing steps aimed at the characteristics of underwater, due to the uniqueness of the underwater imaging environment.The underwater imaging system in neritic regions is easily affected by dynamic fluctuant sunflicker which enormously degrades the quality of the image. These sunflickers in shallow water are produced by the natural light refraction by water ripple. Some special methods are adopted to reduce the influenc e of sunflickers.This thesis focuses on the study of the underwater target recognition problem. Firstly, for the influence of underwater sunflicker, a PCA(Principal Component Analysis) based algorithm has been established. The algorithm is implemented and verified by using VS2010 software and Open CV library. The principle of the algorithm is analyzed. At the same time, Image registration failures are analyzed, and a reliable registration system is built. Then several representative saliency algorithms are compared, and RC(Region Contrast)algorithm is chosen to improve and assist object extraction. A kind of graph-based segmentation method is used to improve the efficiency of RC algorithm. Target texture information and grayscale information are added to RC algorithm, forming a top-down saliency algorithm. The improved RC algorithm is implemented in VS2010. Experimental results show that the improved RC algorithm is more effective in highlighting target. The results of the improved RC algorithm are binarized. Finally, Connected component fast labeling algorithm is applied to binary images. Target templates and connected domains in experimental images are chosen as samples. Six Hu invariant moments of the samples are computed. According to the distributions of invariant moments, the appropriate Hu invariant moments are selected to describe samples and then Bayesian classifier is designed.