Research On 3D Display Algorithm Based On Monocular Endoscope Image

The three-dimensional display of the endoscope can allow doctors to observe the patient’s lesion information and perceive depth more clearly,assist doctors in surgical operations,and thus shorten the operation time.With the rapid development of deep learning,endoscopic image depth estimation has made great progress in accuracy and speed,and gradually replaced traditional methods.In this thesis,the method of deep learning is used to dehaze and estimate the depth of endoscopic images.In this thesis,the Cycle GAN network model is used to purify endoscope smoke images.The generator architecture of the Cycle GAN network consists of multi-scale residual blocks,which help to mitigate the smoke components at different scales,while the adopted refinement module helps to refine and recover more detailed edge structures.Quantitative comparisons with other models are carried out,and the results show that the proposed method has distinct advantages in smoke removal and timeliness of laparoscopic surgical images.When the structural similarity reaches 0.98 and the peak signal-to-noise ratio reaches 32.09,the two indicators indicate that the image after purifying the smoke is similar to the fog-free image,which can restore a more realistic surgical field.The number of frames per second transmitted by the model is about 86.96,which can meet the real-time requirements of the system.Aiming at the problem that the monocular endoscope image lacks the real parallax label,this thesis refers to the methods used in the literature in recent years,takes the result of stereo matching in the traditional method as the real value,and compares the model results in this thesis with the real value.The U-Net network is selected as the basic framework,and the model structure is improved.In order to improve the network performance,the feature maps obtained from each layer in the encoder are output in series.This operation combines semantically rich and spatially accurate features.A lightweight decoder is adopted to lighten the network volume for better system portability,and three different attention mechanisms are explored to improve the accuracy of network predictions.In the laparoscopic dataset,a peak signal-to-noise ratio of 0.8914 and a structural similarity of 16.3614 were achieved.Finally,the obtained depth information is combined with the original image and filled with holes to obtain a right view with a clearer structure.With the help of chromatic or polarized 3D visual aids,the 3D display of monocular endoscopic images is realized.