Light Field Depth Estimation Based On Focal Stack And Improved EPI-Net Network

Light field imaging is an evolving technology for acquiring higher-dimensional visual information.The traditional imaging technology captures the scene from the two-dimensional projection of the light rays of the scene and multiplexes the angular domain,On the other hand,light field imaging captures the luminance information of a same single scene point in different angular directions,thereby the multiplexing of angle information lost in the traditional imaging technology is realized.The use of this high-dimensional representation of visual data allows us to better understand the images of arbitrary scenes and improves the performance of traditional imaging computer vision tasks such as depth estimation,image post-processing,post-Capture refocusing,all-focus images,object segmentation and classification etc.Light field cameras capture the spatial and angular characteristics of light in space.From these features,the depth information of the light field in any lighting environment can be inferred,which is also a huge advantage of active range sensing devices.However,light-field images captured by hand-held light-field cameras contain noise and have very narrow baselines,making depth estimation a challenging task.Researchers have proposed different algorithms to reduce these limitations,However,these methods often require a trade-off between the speed and accuracy of the algorithms.In this paper,depth estimation techniques based on light-field focus stack are investigated.Besides,we propose a fast and accurate light field depth estimation algorithm based on Fully Convolutional Neural Network.In the HCI 4D light field benchmark,we have achieved good performance in most metrics.The specific research work and innovations in this paper are as follows:Firstly,we have proposed a focal stack-based depth estimation method of light field images where we decoded the light field raw images and extracted light field images.After that,create a focal stack to perform digital refocusing and finally obtained final depth map of light field images.This paper investigates the use of decoding and re-sampling methods to Digitally refocus the light field images.To sample the light field,a standard Plenoptic camera(SPC)places a set of micro-lenses in front of the sensor of a conventional camera.In this paper,a closed-form model of the functional relationship between the refocusing distance z and the refocusing parameter ρ of a standard Plenoptic camera(SPC)is presented.The model is based on a first-order optical analysis of the camera imaging process,through which digital refocusing and depth maps of light field images can be obtained.When the traditional algorithm evaluates the image refocusing,there are multiple coordinate base transformations and projection integration operations,which seriously affect the calculation speed of the refocusing image,which in turn affects the acquisition speed of the final depth map.Experimental results show that the use of decoding and re-sampling methods can significantly reduce the time to refocus the light field images without significantly reducing the quality of the acquired depth map.Secondly,we recovered the depth map of light field images using Improved EPI-Net network.When traditional EPI-Net performs depth estimation on light field images,the convolution blocks used to obtain high-level features have problems such as slow speed,many input parameters,inadequate applicable of previous features,low quality and small size depth map etc.Therefore,this paper proposes a depth estimation method for light field images based on an improved EPI network,which combines the initial low-level features and uses dense blocks(Dense-Blocks)to obtain high-level features,thereby reducing the number of parameters while making full applicable of previous features and ultimately improving the quality of the depth map.The experimental results substantiate the practicability and effectiveness of the network.Compared with the traditional EPI-Net network,the light-field image depth map obtained by the improved EPI-Net network has visually richer texture features.The mean square error(MSE)and bad pixels(BP)rates of final depth map achieved through improved EPI-Net are minimized as compared to previously proposed algorithms like that spinner parallelogram operator(SPO)and conventional EPI-Net network...