Research On Learning-based Light Field Image Coding Algorithm

With the development of technology,light field imaging has been widely used in3 d reconstruction,depth estimation,new view synthesis and other fields.Light field imaging is characterized by capturing not only the brightness and color information of the light in the scene,but also the direction information.This means that there are far more data in the light field image than the traditional image,which requires more space to compress and store the light field image and consumes more bandwidth for transmission.The existing coding technology is mainly designed for ordinary images and videos,which can be directly applied to the compression and coding of light field images,but can not effectively remove the information redundancy in light field images.Therefore,it is necessary to design an efficient compression and coding scheme for light field images.In view of the characteristics of light field image,this paper combines deep learning technology with video coding technology to design a new light field image compression coding scheme.As more and more techniques are used in light field image coding,the coding performance is greatly improved.However,the coding complexity is getting higher and higher,and the coding time is longer,which is not good for practical application.When VVC(Versatile Video Coding)is used in light field image Coding,although the performance of VVC is higher than that of other Video encoders,the Coding efficiency is seriously hindered because the Coding unit(CU)partition process consumes a lot of time.In this paper,a fast Coding unit partition algorithm is proposed to shorten the Coding time.Improve coding efficiency.The main research work and innovations of this paper are as follows:(1)Light field Images can be transformed into EPI(Epipolar Plane Images)by re-decomposition and arrangement.EPI is first blurred to remove high-frequency information and prevent aliasing in the following sampling.Then the sampling operation is carried out,and the sampled EPI is restored to the original complete EPI through the convolutional neural network.In this way,only some parameters are encoded and the sampled selected part is encoded into bitstream using VVC.In order to improve the coding performance,VVC coding is used for experimental comparison of pseudo-video sequences.In order to guarantee the image quality of convolutional network,the standard deviation of gaussian blur used was studied experimentally.Experimental results show that the proposed optical field image coding scheme based on EPI recovery network can significantly reduce the bit rate while ensuring the image quality.(2)Most of the current light field image coding algorithms aim at improving the coding performance,without considering the complexity of the algorithm.In this paper,the VVC encoder in the light field image coding algorithm constructed in Chapter 3 is optimized to improve the coding efficiency.In view of the low efficiency of CU partition in VVC coding process,this paper uses XGBoost algorithm to quickly partition CU in VVC frame.Through the statistical analysis of different sizes of CU partition,according to the size of CU partition method.In order to improve the accuracy of XGBoost's classification and discrimination of CU,it is necessary to extract the features in CU,including the size,variance,quantization parameters,horizontal gradient,vertical gradient,maximum gradient and partial gradient difference information obtained in advance.Using the extracted features to make data sets,train XGBoost,integrate the trained XGBoost module into VVC,and realize the rapid CU partition.Experimental results show that the proposed fast intra VVC CU partition algorithm can reduce the coding time by 38.32% on average and BDBR only increased by 0.95% under the all intra VVC coding mode.