In the domain of medical diagnosis and detection, the depth estimation technology of endoscopic images has surpassed the limitations of human vision, offering significant convenience for disease detection and surgical operations through precise depth information. However, in the complex endoscopic environment, challenges such as low illumination, and illumination consistency, resulting in insufficient accuracy of monocular depth estimation. To tackle the aforementioned problems, this paper presents a depth estimation method integrating physical priors and data-driven approaches. This method combines the illumination decomposition method based on the Retinex theory with a self-supervised depth estimation network. Through the illumination-aware feature modulation module, it dynamically enhances the details in low-light regions and suppresses overexposure interference. Additionally, by leveraging the advantages of the self-supervised depth estimation network, a depth estimation network compatible with low-cost and various types of endoscopic devices in the endoscopic environment has been achieved. Experiments on the SCARED and Hamlyn datasets indicate that this method exhibits excellent performance in key metrics.