The quality of images taken in an insufficiently lighting environment is degraded. These images limit the presentation of machine vision technology. To address the issue, many researchers have focused on enhancing low-light images. This paper presents a zero-reference learning method to enhance low-light images. A deep network is built for estimating the illumination component of the low-light image. We use the original image and the derivative graph to define a zero-reference loss function based on illumination constraints and priori conditions. Then the deep network is trained by minimizing the loss function. Final image is obtained according to the Retinex theory. In addition, we use fractional-order mask to preserve image details and naturalness. Experiments on several datasets demonstrate that the proposed algorithm can achieve low-light image enhancement. Experimental results indicate that the superiority of our algorithm over state-of-the-arts algorithms.