Research On Recognition And Location Technique Of Underwater Object Based On Monocular Vision

Autonomous underwater vehicle works as a carrier in complicated marine environment, whose good environmental perception ability serves as important foundation and key technique in improving security and raising its level of intelligence. Robotic vision has a rich sense of information that plays an important role in underwater robot operations and proximity information apperceived. It is of great theoretical significance and of practical application value to conduct research on visual processing technology of underwater robot in specific marine environment.This paper mostly researches the recognition and location of underwater object. Based on the research and of related papers, this paper do research on the methods of underwater image enhancement, underwater image segmentation technology, underwater objects recognition and its location.Aimed at the fuzzy problem in underwater image, An image enhancement method based on minimum cross-entropy is proposed. The method defines a new membership function. The function can't appear the instance of no solution during inverse transformation which ensures the wholes of gray information after image enhancement. The curve shape of function assumes S which can achieve the expectation enhancement effect through the few iterative number of times. The method also combines minimum cross-entropy segmentation algorithm to ascertains the critical threshold value of enhancement and realizes fuzzy enhancement of different image self-adaptivly. Experiments shows that the enhancement effects of the method is better than traditional ones, which enhance the contrast of fuzzy image effectively.Aimed at the problem of illumination inhomogeneous image, the algorithm of segmentation by keeping the intensity moment which is based on dynamic gray-level transform combined with the value of hue is proposed. This method ascertains the transform parameter of dynamic gray-level transform which is used to dynamic gray-level transform with the value of Hue. This method can enhance the part of object and weaken the part of background which is affected by illumination. The algorithm of segmentation by keeping the intensity moment is used to segment the part of background which is effected by illumination. This experiment verifies that this approach is able to realize the segmentation of images effectively in poor distribution of illumination.In the phase of feature extraction, this paper restructures six feature invariant with translating, scale, rotating invariance which improve the scale invariance of Hu's invariant moment. Based on six invariant moment structured in this paper, the conjugate invariant moments of gray and grad is structured which is the input of BP network. BP network is used to recognize underwater objects. The experiments shows the recognition ratio is higher when the conjugate invariant moments of gray and grad is the feature of recognition than when the invariant moments of gray is the feature of recognition.In accordance with underwater specific environment, it uses imaging geometry model combined with second focus method to locate objects position in 3-D and confirms locating parameters with the calibration of camera, whose efficiency are verified through experiments. The location average error is about 10mm.