<p>Low-light image enhancement aims at improving the visual quality of images captured under low-light conditions. However, most of the existing image enhancement methods rely on the guidance of the normal images and suffer from high computational costs. Color and luminance distortions are inevitably introduced when enhancing images, and even ineffective when facing unknown complex scenes in the real world. To overcome the above problems, we propose a self-supervised enhancement method for real world low-light images using Retinex and camera response function. Specifically, we design a lightweight deep network based on Retinex to estimate the illuminance from a low-light image and obtain the exposure ratio map. The correction module is designed based on the camera response function to carefully enhance the illuminance of each pixel, which can effectively reduce the color and light distortion. In order to get rid of the constraints of limited data on the model, the network weights are updated by iteratively minimizing the loss function, so that enhancement of arbitrarily low-light images can be achieved without any training data. Finally, extensive experiments are conducted to provide a comprehensive evaluation of the model. Comparison results on five publicly available datasets with various lighting conditions show that the proposed method presents more natural and pleasing visual effects compared to several state-of-the-art methods. The proposed method only needs 8.9G FLOPs and 13&#xa0;K parameters, achieving a good balance between performance and computational complexity.</p>