Research On Light Field Depth Estimation Algorithm Based On Deep Learning

As an emerging imaging method,light field images can extract multi-dimensional scene information with only one acquisition and can expand two-dimensional images to four-dimensional.Virtual reality,augmented reality,and autonomous driving,which are widely studied now,all need to obtain real depth information in the scene for rendering and calculation.The information contained in the light field has great research significance and research value in the field of depth estimation.The applications of computational imaging are also widespread.Therefore,in order to obtain a more accurate depth map,this paper studies a method for extracting depth information from light field images based on deep learning algorithms.The main research work of this paper is as follows:(1)This paper introduces the development trend of light field theory,several basic acquisition methods of light field images and visualization of light field data.And it also introduces the research status of light field depth estimation,expounds three traditional light field depth estimation methods,and studies the existing deep learning depth estimation methods.And systematically analyze the research status based on traditional algorithms and deep learning algorithms.(2)A light field depth estimation algorithm based on deep learning is designed.In the training process of the neural network,the network parameters are adjusted through the channel attention mechanism so that the calculation target is concentrated on the main objects in the scene,thereby reducing the loss of computing power;at the same time,a feature fusion module is designed for the multi-parameter input of the EPI method.After extracting and filtering the features of the single input,multi-channel fusion is performed,which retains the depth features and textures of the light field image,and improves the learning efficiency and accuracy of the network.A network branch is designed to extract the contour features of the input image and provide edge constraints for the output depth map.The comparative experiments show that this method has obvious advantages over other methods.(3)In order to improve the performance of the neural network,the residual structure is used to increase the depth of the network and improve the accuracy of the depth map output by the network.Based on channel attention,attention in the spatial dimension is added,which makes important features less likely to be lost,and further improves the output performance of the network.