Ghost-Free High Dynamic Range Imaging Based on Two-Stage Dense Image Alignment

High dynamic range (HDR) imaging extends the range of illumination in a digital image beyond its inherent limitations. A common approach to obtaining HDR images is to merge multiple differently exposed images. However, this method can lead to visual distortions, especially in dynamic scenes or when the device used to capture the image is in motion. These distortions manifest themselves as ghosting effects that degrade the overall image quality. In this study, we propose a deep learning-based solution for ghost-free HDR imaging. Our approach uses the RANSAC-Flow algorithm to align low and high-exposure images to a medium-exposure image, gradually refining the alignment from coarse to fine. We further improve the process by using a specialized merging network inspired from Exposure Structure Blending Networks (ESBNs). Our merging algorithm takes aligned images along with low, medium, and high exposure images as input and uses a convolutional neural network (CNN) to generate HDR images. To evaluate our method, we performed experiments by merging datasets from SIG17 and ICCP19. Quantitative experiments show a significant improvement over the state-of-the-art method, with an increase of 3.7 dB in the Peak Signal to Noise Ratio (PSNR) and an increase of 0.07 in the Structural Similarity Index Measurement (SSIM) metric. Furthermore, qualitative experiments on the validation sets of the combined datasets show the effectiveness of our proposed method in suppressing ghosting effects while improving image details, resulting in visually pleasing HDR images.