Self-supervised Binocular Vision Based On Neural Network And Its Application In Blood Vessel Depth Measurement

In recent years,with the development of interdisciplinary research,the research on stereo matching theory has gradually been enriched.However,medical images are generally characterized by high noise,complex structures,and difficult data acquisition.There are problems in traditional binocular matching techniques to obtain accurate matching results.Besides,it is difficult to obtain accurate label data,which restricts supervised training.Aiming at the above problems,this paper introduces the self-supervised method to the stereo matching of medical images and uses deep learning to improve the matching accuracy.The main research work is as follows:(1)Aiming at solving the problem that it is difficult to obtain label data,a self-supervised image stereo matching model SDMNet(Self-supervised Disparity Matching Network)is proposed.This model introduces four modules: self-supervision,perceived loss,region cropping,and loss equalization calculation.The model achieves the highest accuracy of mainstream self-supervised algorithms on the KITTI public data sets,and it obtains 3.39%and 2.25% error rates on KITTI 2015 and 2012 data respectively.(2)Aiming at solving problem of occluding points in stereo matching,based on SDMNet,a sparse auto-encoder image stereo matching model ORB-SDMNet is constructed by using a priori driven data learning,fast feature point extraction and description algorithm(Oriented FAST and Rotated BRIEF,ORB).This model makes full use of geometric priors,introduces three modules: sparse point cloud generation,sparse auto-encoder,and loss equilibrium calculation.It is compared with other algorithms on the public data sets,surpasses some supervised algorithms,and obtains 3.31% and 2.20% error rates on KITTI 2015 and 2012 data respectively.(3)According to the demand of three-dimensional imaging of vascular imaging and pathological diagnosis,combined with the image data characteristics of blood vessels and related tissues,a non-invasive high-precision three-dimensional imaging technology of blood vessels based on the two image stereo matching models established in this paper is proposed,and a three-dimensional vascular imaging system is built.The system can output satisfactory qualitative imaging results.To simulate the imaging characteristics of the back of the hand and the dorsal blood vessel imaging environment,a three-dimensional simulated blood vessel imaging system was built.By collecting simulated blood vessel data and obtaining disparity labels,the self-supervised stereo matching model algorithm was quantitatively evaluated on the simulated data.It achieves an effect of 0.57 MAE(Mean absolute error)error lower than mainstream algorithms.