EPI-based Light Field Depth Estimation And Angle Reconstruction

The light field is an important way to record spatial scene information.Recording three-dimensional spatial scenes from different perspectives can digitally store the target area information.It plays an irreplaceable role in the fields of three-dimensional reconstruction,digital refocus,and so on.Due to the high coupling between light angle information and scene depth information,extracting scene depth information from light field data has become the focus of light field theory research.Depth information can be used in intelligent vision technologies such as image super-resolution reconstruction,robot vision,and digital twin,which brings new opportunities for the development of intelligent vision.Based on the EPI information and structure of light field image,this paper designs and constructs the light field depth estimation framework EAnet and the light field consistency angle resolution reconstruction method.The main contents are as follows:To solve the problem of ignoring the spatial characteristics of different viewing angles and large error of depth estimation details in the process of depth estimation,a depth estimation network EAnet based on multi-stream EPI feature extraction and attention model prediction is designed in this paper.Firstly,according to the angle characteristics of light field data and the principle of EPI prediction depth,the multistream network is designed to preprocess the input image and extract the features;Then,the attention model is designed to predict the light field image and form the attention map with assigned weight,to improve the efficiency of data utilization;Finally,the multi-stream network features are combined with the attention map to estimate and optimize the overall depth,and the final depth estimation results are obtained.Experiments are carried out on HCI data sets.Compared with a variety of classical methods,this method is greatly improved.At the same time,ablation experiments are designed,and then the improvement of depth estimation results by this method is analyzed.To reduce the aliasing artifacts in the reconstruction of light field angular resolution and improve the detail quality of reconstruction results,a light field reconstruction method based on consistency analysis optimization is designed in this paper.Because the direct reconstruction of the full image will cause the ringing effect,the detail reconstruction is selected on the EPI slice of the light field image.Firstly,the light field data is stacked and sliced into horizontal and vertical EPI information,and then the original EPI information is sampled to a lower resolution;The convolution neural network is used to learn and reconstruct the low-frequency EPI information features,and the low-frequency EPI features are restored to the target angle resolution according to the sampling kernel function to obtain the target angle reconstructed light field image;The consistency of the reconstructed image is analyzed,and the details of the low-quality results are restored and reconstructed again.By using the consistency results to adjust the neural network learning process,the reconstructed image quality is improved.Experiments on HCI and Stanford data sets show that this method is superior to the current mainstream methods in both subjective and objective experiments.