Research On High Dynamic Range Image Reconstruction And Display Based On Deep Learning

High Dynamic Range(HDR)images can accurately represent the brightness range from sunlight to the darkest shadow in the real scene.HDR images have a larger color and brightness than Low Dynamic Range(LDR)images.Recent years,with the development of ultra-high-definition video services and HDR technology,the demand for HDR images has gradually increased.The conventional way to acquire HDR images is to shoot directly with an HDR camera.However,such cameras are very expensive.Another approach is to reconstruct HDR images from the LDR images obtained by conventional acquisition equipment through some algorithm,recovering the lost details in over-exposure and under-exposure regions.In addition,since the dynamic range of HDR images is larger than that of images that cannot effectively display HDR content by ordinary monitors,and a display device that supports HDR content is required to make the displayed content closer to the real scene.Therefore,how to better reconstruct HDR images from LDR images and improve the display effect of HDR images has important research significance.Based on this,the research work of this thesis is as follows:1.Due to the limitations of the image sensor,the details are lost in over-exposure and under-exposure regions.It is very challenging to reconstruct HDR images directly from a single LDR image.To reconstruct high-quality HDR images,this thesis proposes an HDR image reconstruction method based on a dual-branch cascade network,which includes a multi-exposure LDR image generation stage and an HDR image reconstruction stage.In the multi-exposure LDR image generation stage,a dual-branch cascaded network is proposed,each branch is cascaded from high-exposure or lowexposure networks with the same network structure to generate multi-exposure LDR images.In the HDR image reconstruction stage,the multi-exposure fusion method is used to merge the multi-exposure LDR images generated by the network.The experimental results show that the method can more effectively restore the lost details in over-exposure and under-exposure regions and reconstruct high-quality HDR images.2.In the multi-exposure fusion method based on the radiation domain,the inverse camera response function estimated by the LDR image is different from the inverse camera response function obtained by the real camera response function,the mapping relationship between the multi-exposure LDR image and the HDR image cannot be effectively established,the effect of the fused HDR image is poor.To improve the quality of the fused HDR image,this thesis proposes an HDR image reconstruction method based on a multi-exposure fusion network is proposed.The proposed method mainly uses the multi-exposure LDR image as input and reconstructs the HDR image through the multi-exposure fusion network.Two LDR images with adjacent exposure times are fused using a multi-exposure fusion sub-network,and a feature map is obtained after calculation.Then,the feature maps obtained by each exposure fusion sub-network are concatenated and then go through a convolutional layer to predict the final HDR image.The experimental results show that the proposed method can effectively improve the quality of the fused HDR image.3.The traditional local dimming method is difficult to apply to the display of different types of HDR images.To obtain the backlight luminance and LCD panel display image that improve the display effect of HDR image,this thesis proposes an HDR image local dimming display method based on deep learning.First,combined with the principle of the conventional local dimming method,the backlight luminance extraction network and the LCD panel display image generation network are proposed.Then,the backlight luminance dataset and LCD panel display image dataset to improve the display effect of HDR images are constructed by the subjective and objective methods.Finally,the backlight luminance extraction network and LCD panel display image generation network are trained separately based on the constructed dataset.The experimental results show that the proposed method can not only reduce energy consumption but also more effectively improve the display effect of HDR images.