| With the continuous development of medical science and technology,the accuracy requirements for single hole minimally invasive surgical robot endoscopic imaging systems are constantly improving.Traditional endoscopic imaging systems capture two-dimensional images,which lack necessary depth information,increasing the difficulty and risk of surgery.However,stereo vision systems based on binocular endoscopes can obtain the three-dimensional coordinates of lesion points,which can better ensure the safety and stability of surgery.Therefore,this article takes binocular endoscopes of the single-hole minimally invasive surgical robotic as the research object,and conducts a detailed study of the specific process of binocular stereo vision.Commonly used binocular cameras are used for experimental verification,focusing on improving the accuracy of traditional stereo matching algorithms and the balance between the prediction accuracy and speed of stereo matching networks.The improved algorithm is applied to 3D reconstruction of real scenes,and the reconstructed results are compared and analyzed.Firstly,the AD-Census stereo matching algorithm is improved.A gradient transform based on the cross arm is proposed,and it is fused with TAD transform and Census transform to form a new matching cost calculation method.The improved method is applied to the classic stereo matching test map of Middlebury for testing.The test results show that compared with the unimproved algorithm,the average error matching rate of the proposed stereo matching method has decreased by 12.0%;The stereo matching accuracy on Venus images has been improved by 56.5%,which greatly improves the problem of poor matching accuracy in occluded areas and areas with discontinuous disparity using traditional AD-Census stereo matching algorithms.Secondly,a stereo matching network based on multi-level cascaded recurrent is proposed.A multi-layer network is designed that introduces position encoding and self-attention mechanisms on high-resolution feature maps,and utilizes disparity refinement strategies such as hierarchical cyclic refinement,cascading refinement,and cyclic refinement to update disparity values,to better restore the detailed information of stereo images.In addition,the strategy of disparity iterative update has been improved by using lightweight group correlation layers at low scales and adaptive group correlation layers at high scales to update the difference values,to reduce the computational complexity of disparity iterative update.Designed experimental verification schemes on multiple stereo matching datasets and conducted comparative analysis.The experimental results show that under the same dataset and training parameter settings,the proposed algorithm improves disparity prediction accuracy by 19.6%and model inference speed by 46% compared to the cascaded cyclic stereo matching algorithm;On the Middlebury stereoscopic evaluation platform,the proposed algorithm has a relatively fast model inference speed compared to other algorithms,while achieving highly competitive disparity estimation accuracy.In addition,the proposed stereo matching algorithm achieves a good balance between the accuracy of disparity estimation and the speed of model prediction.Finally,to compare and analyze the matching effect of the proposed stereo matching algorithm in real scenes and its impact on the accuracy of point clouds in 3D reconstruction,commonly used binocular cameras and binocular endoscopic imaging system of the single-hole minimally invasive surgical robot are used to collect left and right images in real scenes,and stereo correction,stereo matching,and 3D reconstruction experiments were conducted sequentially.The experimental results show that compared to the improved traditional stereo matching algorithm,the proposed stereo matching network has higher stereo matching accuracy in real scenes,higher 3D reconstruction point cloud accuracy,and can better reconstruct the overall contour of the target image. |
PDF Full Text Request