HDR Image Reconstruction From A Single LDR Image Using Deep CNN

High dynamic range(HDR)imaging is a rapidly growing field in image processing and computer graphics.It is a photographic technique that can represent a wide range of real-world luminosity.In contrast,conventional imaging also known as low dynamic range(LDR)imaging is not capable of representing a real-world range of intensities and colors since most digital cameras can capture a limited range of light intensity in a natural scene.HDR imaging overcomes such limitations allowing capturing and delivering information that can match the dynamic range of the human visual system.This HDR information can be used in various applications,such as tone-mapping,image-based lighting,HDR displays,computer vision,and other post-processing operations.HDR imaging has seen a lot of development but despite this,LDR imaging remains the standard today in terms of content.To use this content in different HDR applications,LDR content needs to be converted and expanded to HDR where missing information lost due to saturation or quantization needs to be recovered.The goal of this thesis work is to develop deep neural networks for the purpose of reconstructing an HDR image from a single exposure LDR image.We utilize convolutional neural networks capabilities to overcome the limitations of the previous works and deal with the problems encountered when reconstructing HDR content such as overexposed areas,under-exposed areas,quantization,and linearization.In this work,we propose two novel CNNs for HDR image reconstruction.The networks are trained on a large dataset of HDR images using recent advances in deep learning.The first method uses the deep supervision technique on a convolutional auto-encoder network.Deep supervision adds companion loss functions at each hidden layer in addition to the overall loss function at the output layer.This helps the network to easily recover structure in lower resolution and recover texture in higher resolution.The second method adopts a convolutional auto-encoder that uses residual learning techniques to accelerate training,attention mechanism to enhance useful features,and contextual attention module to help recover information lost in over-exposed areas.Extensive quantitative and qualitative experiments on public HDR datasets demonstrate the ability of our proposed methods to effectively reconstruct visually pleasing HDR images and show better performance compared to existing approaches.