Research On Component Recognition And Ranging Of Underwater Christmas Tree Based On Binocular Stereo Vision

The global demand for oil resources is increasing recently,and offshore oil and gas development has become one of the directions that people pay attention to.Due to the complex environment of underwater deep-sea oil and gas development,ROV plays an irreplaceable role in the installation and maintenance of subsea trees.However,in the existing underwater operations,it is difficult for the operator to intuitively judge the position and type of components only by relying on the two-dimensional video sent back by the camera equipment.In order to obtain the depth information of the Christmas tree components in the camera picture,a binocular stereo vision algorithm is required to match the left and right camera pictures.Therefore,this thesis uses a semantic segmentation algorithm to first perform semantic segmentation on underwater images,and then only match the segmented areas,which overcomes the problem of object edges in underwater stereo matching to a certain extent.Blur problems and weak texture problems,so as to obtain more accurate depth information.This thesis mainly researches in the following aspects.1.First,the underwater image enhancement is studied.Due to the environmental conditions of underwater imaging,the resulting image data quality is much lower compared to land images.Facing the low-quality underwater images caused by the absorption and scattering of light by water bodies,this thesis studies and analyzes the existing common underwater image enhancement methods.Through the research on the principle of the method and the comparison of the imaging effects,this thesis improves the dark channel prior dehazing algorithm,and proposes an underwater image enhancement algorithm combining the inverted dark channel prior dehazing algorithm and the gamma correction method.On the basis of restoring the shape of the target object in the water,the brightness balance of all parts of the screen is guaranteed,which establishes the foundation for the subsequent segmentation tasks.2.Facing the task of detection and segmentation of subsea tree components,the structure analysis of Deeplab V3+ network is carried out.Due to the large amount of computation in the backbone feature extraction network of the existing network,this thesis introduces the IRS inverted residual module,which first performs the dimension-up operation on the input to refine the features,and then performs the dimension-reduction operation to ensure that each output channel can contain more dimensions.Characteristics.The number of parameters of the hole convolution module of the existing network is large,and the setting of the expansion coefficient will lead to the problem of holes in the receptive field.The structure of the existing ASPP module is improved and the parameters are adjusted,and the dimension of the module is reduced while expanding part of the parallel connection.The convolution module is changed to a series structure,which reduces the amount of calculation and increases the receptive field of the network by stacking and expanding the convolution module.By introducing the attention mechanism module,the weights of different feature layers are adaptively adjusted,so that the feature layers that have a greater effect on the result output have greater weights.Finally,through the comparison of ablation experiments,the effectiveness of the network improvement in this thesis is verified.3.For the last step of binocular stereo matching,the traditional stereo matching method is studied first,and the semi-global matching algorithm is selected according to the working conditions.By introducing a dynamic disparity array,the algorithm can use different disparity ranges for objects in different pictures,which improves the accuracy and continuity of the matching results of depth discontinuity objects.At the same time,by improving the sliding window structure of the SAD step,the window resolution is maintained.The receptive field of the sliding window is expanded when the situation changes,thereby improving the accuracy of the algorithm results while ensuring the algorithm operation speed.In this way,the image input to the underwater tree recognition and ranging algorithm based on binocular stereo vision is firstly corrected by the epipolar line,and then input to the semantic segmentation network after image enhancement,and the segmentation result containing only part of the image of the target is output,and then input to the improved model in this thesis.The semi-global stereo matching algorithm performs stereo matching only in the target area,further overcoming the problem of blurred object edges in underwater stereo matching.Finally,the effectiveness of the algorithm in this thesis is verified through qualitative and quantitative comparisons.