Research On High Dynamic Range Video Inverse Tone Mapping Algorithm Based On Deep Learning

In recent years,video services has developed rapidly,and the number of users has also increased.The increasing user demand for video services has promoted the development of the next generation of high-quality video standards and video technologies.Ultra High Definition(UHD)video,especially the High Dynamic Range(HDR)video,is one of the most important directions in recent years.With the maturity of HDR display technique and HDR transmission technique,the difficulty in HDR production and the lack of HDR resources have limited the further development of HDR technology.Based on the latest HDR standards,technical parameters,and general production processes,we put our attention on the inverse tone mapping methods.Inverse tone mapping is a technology which is used to convert a Standard Dynamic Range(SDR)image or video to HDR.Because inverse tone mapping methods take SDR images and videos as input,it can make full use of the existing plenty of SDR images and videos to pile up HDR resources.Inverse tone mapping methods can also be used for photography to use the SDR photographic equipment to get HDR images and videos.Inverse tone mapping provides an efficient and convenient HDR production method.With the development of deep learning,using deep learning to solve inverse tone mapping problems has been a hot research topic.Compared with previous methods,inverse tone mapping methods based on deep learning have greatly improved the effect and versatility,but the existing methods still have problems.Existing methods are usually designed for images.When they are applied to video,flickering usually occurs.Then,they are usually designed for normal images,and cannot be applied to overexposed and underexposed images.At last,they suffer from color distortion.To reduce the flickering in video inverse tone mapping,we propose a video inverse tone mapping method based on 3D convolutional neural network,and propose a dataset production method based on HDR10 video.In training process,the data pre-processing method for inverse tone mapping problem is used,and a loss function including MSE loss,intrinsic loss,and perceptual loss is used.The proposed method improve the visual effects and reduce the flickering.Compared with other inverse tone mapping methods,the proposed method has improved visual effects and has higher scores on objective evaluation metrics.In ablation experiments,the loss function and data pre-processing method have been proved to be effective.For the problem of different exposed inputs and color distortion,we have considered the feature of the video,the theoretical basis of luminance and color,luminance equalization,and color correction.Based on the previous proposed video inverse tone mapping method,we further propose a video inverse tone mapping method using luminance equalization,and a dataset production method based on color correction.We process the luminance and color information separately in (4 (1 color space and use two 3D convolutional neural networks to convert the luminance and color respectively.Compared with other inverse tone mapping methods,the proposed method is capable of processing inputs with different exposures,and has improved objective evaluation metrics and color accuracy.In ablation experiments,we verified the effectiveness of the data preprocessing method and the luminance and color processing method.