Ambient occlusion has been widely used in many graphics and vision tasks for approxi-mating low frequency global illumination, removing inter-reflection, and inferring scene geometry. Existing methods need either scene geometry or scene appearances under different lightings to compute the ambient occlusion of the scene, which limits the application of these methods in many real applications. In this paper, we present a new method for computing the ambient occlusion from a single image taken under unknown natural lighting. Our method needn’t any scene geometry or illumination information. To this end, we utilize a convolutional neural network (CNN), trained on a large amount of synthetic data, to directly recovery the pre-pixel ambient occlusion. We propose three network structures for ambient occlusion estimation, a cascade one and a parallel one that are based on previous CNN solution for intrinsic images, as well as an end-to-end neural network. We analyze their performance and demonstrate that comparing with parallel and cascade designs, the end-to-end design could achieve the best performance. We also valid the efficiency and effectiveness of our method on both synthetic and real images. It is not only faster, but also more accurate than previous methods.