Recovering a high dynamic range (HDR) image from a single low dynamic range(LDR) input image is challenging due to missing details in under-/over-exposedregions caused by quantization and saturation of camera sensors. In contrast toexisting learning-based methods, our core idea is to incorporate the domainknowledge of the LDR image formation pipeline into our model. We model theHDRto-LDR image formation pipeline as the (1) dynamic range clipping, (2)non-linear mapping from a camera response function, and (3) quantization. Wethen propose to learn three specialized CNNs to reverse these steps. Bydecomposing the problem into specific sub-tasks, we impose effective physicalconstraints to facilitate the training of individual sub-networks. Finally, wejointly fine-tune the entire model end-to-end to reduce error accumulation.With extensive quantitative and qualitative experiments on diverse imagedatasets, we demonstrate that the proposed method performs favorably againststate-of-the-art single-image HDR reconstruction algorithms.