How to obtain object information as rich as possible, with the highest possible speed and accuracy from recorded optical signals, has been a crucial issue to the pursuit of powerful imaging technologies. Nowadays, the speed of ultra-fast photography can exceed one quadrillion. However, it can record only two-dimensional images which lack the depth information, greatly limiting our ability to perceive and to understand the complex real-world objects. Inspired by recent successes of deep learning methods in computer vision, we present a novel high-speed three-dimensional (3D) surface imaging approach named micro deep learning profilometry (μDLP) using the structured light illumination. With a properly trained deep neural network, the phase information is predicted from a single fringe image and then can be converted into the 3D shape. Our experiments demonstrate that μDLP can faithfully retrieve the geometry of dynamic objects at 20,000 frames per second. Moreover, comparative results show that μDLP has superior performance in terms of the phase accuracy, reconstruction efficiency, and the ease of implementation over widely used Fourier-transform-based fast 3D imaging techniques, verifying that μDLP is a powerful high-speed 3D surface imaging approach.