DCDR-UNet: Deformable Convolution Based Detail Restoration via U-shape Network for Single Image HDR Reconstruction

Joonsoo Kim, Zhe Zhu, Tien Bau, Chenguang Liu; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 5909-5918

Single image based HDR reconstruction methods using deep neural network have been proposed to mainly restore the lost details in the overexposed region. However they cannot restore the details well if the overexposed region becomes large because the receptive fields of their networks are not large enough to cover the region. Also they cannot restore the partially overexposed small object well if the non-overexposed portions of the object are sparse. In this paper we propose new deep neural network namely DCDR-UNet (Deformable Convolution Based Detail restoration via U-shape network) for single image HDR reconstruction. By introducing a new block called Deformable Convolution Residual Block (DCRB) and our loss function we show how deformable convolution can be well utilized to solve the problems of the existing methods in single image HDR reconstruction. Our experimental results show that our method achieves much better results than all the existing methods quantitatively and qualitatively.