Images captured under complex lighting conditions often suffer from mixed exposure problem, where both overexposed and underexposed regions coexist within the same image. Due to the non-uniform distribution of mixed exposure, existing methods often struggle to restore both color and details. To address this challenge, we propose a multi-stage reflectance-guided exposure correction (MREC) framework, which consists of three modules: preliminary enhancement, adjustment, and fusion. In the preliminary enhancement module, we design a bidirectional illumination estimation module (BIEM) that takes paired mixed/normal exposure images as input. Through bidirectional illumination estimation, we obtain both a brightened version and a darkened version of the mixed exposure image as its preliminarily enhancements. Meanwhile, we derive the forward and inverse reflectance images from the normal exposure image. In the adjustment module, guided by the reflectance images as supervision signals, we further optimize the preliminarily enhanced images in the wavelet domain by introducing the low-frequency diffusion restoration (LDR) and the high-frequency refinement (HFR). In the multi-scale non-local fusion module (MNFM), we integrate the adjusted images along with the original mixed exposure input, producing an output with globally well-balanced exposure. Experiments demonstrate that our method produces results with satisfactory perceptual fidelity, and achieve competitive performance in terms of distortion and perceptual metrics.