A Research On Image Feature Extraction And Recognition Based On Neural Networks Of Underwater Object

The subject stems from the project of "information processing and understanding technology of underwater optical vision" in military intelligent underwater robot technology. It is important to smoothly perform the tasks of detecting the underwater objects and getting their position in the complex condition of the underwater.The work of thesis is mainly concentrated on the geometric primitive extraction, feature extraction and recognition in the real underwater environment. Due to the characteristics of the underwater image, the ATR(automatic target recognition) included image collection, image processing, feature extraction, pattern recognition .Combined with the edge detection of dynamic threshold, the improved RHT increases the precision of the boundary point detection, especially for the uneven illumination of the underwater image. In the improved RHT, gradient direction information, of which precision is not required to be high , is used to determine whether the parameter should be accumulated or not. In comparison with the basic RHT, the problem of useless accumulations is well solved.After the feature extraction and selection, the eigenvectors with better clustering effect are got. The influence of scale factor on moment invariant features in discrete are considered, six moment invariants with scale, translating and rotating invariance are proposed. Adding the corner and pad , We can get eight feature vectors. Improved Back_Propagation Neural Net Method is applied to recognize underwater object. To make the BP efficient, adaptive learning speed and momentum are introduced. Theoretical and experimental results show that the precision of recognition is high.