This paper presents a method of deep learning-based interference-free hologram generation. In the method, simulated off-axis digital Fresnel holograms are utilized as network training samples, and an improved convolutional neural network is used to learn the feature relationships of the zero order with the positive and negative first orders of the holographic spectra. The negative first-order spectra of the holograms are thereby extracted. Experimental verification is carried out with simulated holograms and real ones, and the reconstructed images of the interference-free holograms are analyzed. The results show that the proposed method can eliminate zero-order information and interference information in a wide range in the absence of manual intervention, extract negative first-order information from the hologram, and obtain an object light field with high reconstruction quality. This means that the proposed method achieves deep learning-based interference-free hologram generation.