Research On Image Reflection Interference Removal Technology And Its Application In 3D Reconstruction

With the increasing demand for imaging quality of camera equipment,the removal of reflected noise has become a research hotspot in the field of computer vision.The reflection interference will not only affect the visual effect of human eyes,but also bring difficulties for advanced image processing such as depth estimation,three dimensional(3D)reconstruction and so on.At present,it is convenient and efficient to eliminate the reflection interference in the picture with the help of the deep learning tool,however,due to the complexity of reflection,this task still faces many problems,such as the limited application of the algorithm,the serious loss of high-frequency details after the removal of reflection interference,and so on.In order to solve these problems,this thesis studies the reflection interference removal algorithm and its application in 3D reconstruction.The specific research contents of the thesis are as follows:(1)Starting from the optical theory of mixed image formation,a mathematical model of polarized mixed image is established.Based on this model,this thesis deeply analyzes the factors that affect the formation of mixed images,so as to realize the synthesis of mixed images that fit the natural image and establish a polarization image dataset.By observing the typical situation of reflection in real life,we firstly preprocess the reflected and transmitted components of the mixed image.And then the dynamic scenes that may appear in the reflection layer and the incidence angle of light wave are reasonably simulated.At the same time,it is found that the value of the orientation of the plane of incidence has an important influence on the polarization mixed imaging,and then,according to its characteristics,a diversified dataset of light intensity variation trends is established.(2)In order to improve the quality of the output transmission layer after reflection interference removal,this thesis studies a polarized image reflection interference removal network based on attention mechanism which pays attention to the channel and spatial features with greater contribution at the same time.Based on the encoder-decoder network integrating convolutional block attention module,a two-level structure is constructed.After the reflection layer is predicted,the transmission layer is further predicted with the auxiliary information of the reflection layer,and then a more powerful end-to-end polarization image reflection interference removal network is implemented.(3)The research on the application of reflection interference removal algorithm to3 D reconstruction is carried out.On the basis of deeply analyzing the influence of reflection interference on the quality of image 3D reconstruction,the depth map before and after the elimination of reflection interference is obtained by using the depth estimation method based on hybrid vision Transformer,and then the depth map is converted into 3D point cloud by using coordinate transformation relationship.(4)The feasibility of the reflection interference removal algorithm in this thesis is verified by comparing it with existing algorithms from qualitative and quantitative dimensions.The results show that the proposed algorithm can generate transmission image with a good visual effect,and the peak signal to noise ratio and structure similarity index of the transmission image can reach 32.43 d B and 0.960.By comparing the depth map and 3D reconstruction result before and after reflection interference removal,the feasibility of the reflection interference removal algorithm applied to 3D reconstruction is verified.