Digital lithography based on digital micromirror array devices is one of the methods for the production of micro and nano structures, with the advantages of low cost and high flexibility, and has great advantages in micro and nano processing. As the demand for higher resolution of micro and nano structures increases, the wavelengths of lithography systems are getting shorter and the numerical apertures are getting larger, which leads to shorter and shorter exposure depths of focus. To ensure the quality of lithography patterns, the substrate must be within the depth of focus, so fast and high precision inspection of the focal plane becomes the key to the production of high resolution micro and nano structures. Most of the traditional methods require a separate design of the focus detection system, which will not only increase the complexity of the whole system structure but also increase the difficulty of mounting. With the growing development of image processing technology, focusing methods do not require complex optical path adjustment and can achieve focal plane detection based on out-of-focus images only, and the methods have been applied to many fields. Inspired by this, this paper proposes a deep learning based focus detection method, by adjusting the optical path of digital lithography so that the exposure focal plane and the camera imaging focal plane coincide, at this time the image captured by the CCD is the exposure pattern on the exposure focal plane, the blurred degree of the image directly reflects the out-of-focus degree of the exposure pattern, so the focal plane can be quickly and automatically detected using the algorithm only based on the currently formed image. The focus detection algorithm proposed in this study consists of two steps, firstly a coarse focus detection of the substrate at a large out-of-focus distance to reduce the out-of-focus range of the substrate, and then a further improvement of the focus detection accuracy at a small out-of-focus distance. According to the characteristics of these two focusing steps, a deep learning model is used for detection in the coarse focusing and a conventional sharpness evaluation function combined with a search algorithm is used for detection in the precise focusing. Different out-of-focus ranges are firstly classified according to the out-of-focus distance, and corresponding training and test datasets are produced. The trained network can achieve 88.7% accuracy on the test set, and it only takes 90 ms to determine the current out-of-focus range of the substrate, and then move the displacement table to move the substrate to the focal plane. Compared to conventional methods, this avoids the need for a round trip movement of the displacement stage, thus reducing the impact of return errors on focus detection accuracy. The evaluation performance and evaluation speed of different sharpness evaluation functions were also compared using out-of-focus image data. The Laplacian function is chosen as the image sharpness evaluation function for precision focus detection. Using this function, the sharpness value of an image can be calculated in only 5 ms and combined with the search algorithm to find the position with the highest sharpness value near the focal plane, the focal plane can be found accurately in 7 steps on the basis of coarse focus detection. Simulation and experimental validation results show that although the method has a certain chance of error in the coarse focusing stage, the misjudged out-of-focus range is small and can be corrected by the search algorithm in the precise focusing stage to find the true focal plane. In the end, the method can be used with a 5× focusing objective and the focusing accuracy can reach 2 μm in the out-of-focus range of (-40 μm, 40 μm) with a total time of less than 300 ms. In summary, the method has the advantages of low structural complexity, fast focusing speed, and high focusing accuracy, can be well applied in the field of digital lithography.