IID-MEF: A multi-exposure fusion network based on intrinsic image decomposition

This paper follows the idea of divide and conquer to propose a multi-exposure fusion network for the unsupervised generation of high dynamic range-like images. We develop a new intrinsic image decomposition (IID) network based on our modified IID model to produce the reflectance, shading, and color components from source images, which respectively represent the texture structure, lighting condition, and visible color distribution of the imaging scene. Three fusion sub-networks are then designed to process these different components, producing pleasing images with rich structures, reasonable lighting, and suitable color. Specifically, we first develop a reflectance fusion network to integrate the reflectance components, in which an adaptive pixel-scale selection strategy is adopted to dynamically guide the network to preserve rich scene textures. Then, a shading fusion network with the adaptive global image-scale weighting strategy is designed to fulfill lighting adjustment, which can generate suitable illumination by defining the lighting adjustment as a game between the illumination of source images. Finally, we propose a color fusion network that embeds both pixel-scale and image-scale strategies, which can guarantee a pleasing scene color distribution with a specifically designed color fidelity loss. Extensive experiments demonstrate the superiority of our method over state-of-the-art methods in terms of texture integration, illumination adjustment, and color fidelity. Moreover, our method can be applied to fulfill tasks of the single low-light image enhancement and the single overexposed image correction with promising performance. The code is publicly available at https://github.com/HaoZhang1018/IID-MEF.