DBHDR: Dual branch Network Guided Multi-Exposure HDR Image Reconstruction
Ziwei Pang, Huihui Bai · 2022
This paper addresses the problems of artifacts, blur, and color patches that can result from multi-exposure fusion when there are more pronounced moving objects in multi-frame LDR images. We propose a new dual-branch high dynamic range image reconstruction network, namely the multi-frame images fusion branch and the reference image structure enhancement branch. In the fusion branch, we perform deformable convolution between the reference feature and adjacent features to perform feature alignment, and increase the weight of detailed features such as texture in the high-light and low-light regions. The aligned features are concatenated and fed into the dilated residual dense blocks (DRDBs) to obtain dense local features. In the enhancement branch, we use six cascading residual local feature blocks (RLFBs) modules to extract structural features, and then add the shallow features of the input reference image to the extracted features to ensure that the network can guarantee the structural information of the reference image. Through the comparison of experimental results, it is verified that the proposed algorithm can remove artifacts and color blocks while maintaining image saturation and detail recovery in the shadow area.