Research On Registration Of CT Volume Data Based On Feature Point

As an advanced means of industrial Non-Destructive Testing(NDT), Cone Beam Computed Tomography(CBCT) can obtain high-resolution 3D image, i.e. the CT volume data, which provides the details of the structure and the shape inside the interested object, from projections. In practical applications, limited by the view of CBCT system, in order to get fine structural information inside the sophisticated large-size-objects, the common practice is to first image every sub-regions of the object separately. And then all the CT volume data of different sub-regions are mosaiced into a whole one with high resolution. As the key technology of CT volume data mosaicing, registration of CT volume data determines the accuracy of the mosaic result. Therefore, research on registration of CT volume data is of great significance in improving the quality of CT volume data and the accuracy of NDT.In order to mosaic the CT volume data, this thesis focuses mainly on registration technology of volume data based on the feature point,including 3D feature point detection, 3D feature point description and matching. The main works are presented as follows:Aiming at the problem of large-computing and time-consuming in multi-scale feature point detection for 3D volume data, a new method called center surround feature point detection is proposed based on the 2D CenSurE detector. Combined with the 3D integral image, this method can compute the feature response for one voxel very quickly, so the efficiency of building the scale space of 3D volume data is improved obviously. Meanwhile, based on the conventional edge suppression, Harris criterion has also been used to remove the points with high edge response, further enhancing the stability of feature point. The experimental results show that the time spending on our method is reduced significantly and the repetition rate is increased to some extent.Aiming at the mismatching problem caused by similar features in the feature point matching, we propose a new feature point matching method for volume data based on 3D binary descriptor. This method firstly extends the original BRIEF descriptor to the 3D case and then two kinds of binary descriptors with different sampling geometries are extracted in the neighboorhood of the feature point, which has a fixed scale size, to suppress the local ambiguity of the 3D extended feature descriptor. Finally, a coarse-to-?ne feature matching method is designed according to the characteristics of feature descriptor mentioned above. In addition, we also estimate the orientation for each feature point, computing the local gradient of point pairs sampled in the regular pattern and then execute the orientation standardization for each feature point. The experimental results show that the proposed method brings down the number of mismatch, and improves the efficiency of matching obviously. At the meantime, our feature point descriptor has good rotation invariance.Aiming at the problem of the inefficiency in the registration of CT volume data, parallel computing technique implemented on GPU is used to accelerate the process of image registration. Based on the full analysis of the parallelism of feature point detection, we design the parallel program of the presented method by assigning threads configuration and using texture storage legitimately. Experimental results show that the acceleration ratio can reach above 20 with GPU. And after acceleration the feature detector can satisfy the needs of our practical application.