Typically, star images captured in the atmosphere during daylight hours have a strong background and low signal-to-noise ratio (SNR), which makes it difficult for traditional algorithms to extract the star from the images. To improve the recognition rate, we propose an accurate method for simulating star images and train a deep convolutional neural network with a downsampling layer using the simulated images. The trained network can denoise and enhance the star images. Experimental results demonstrate that the proposed method improves the peak SNR by 11.28 dB within an average runtime of 0.2 s, which is significantly less than that of a traditional neural network. In addition, we test the proposed method on the trained network using real star images and find that the improved SNR is 88.9 times greater than that of the existing methods.