An optical watermarking reconstruction method based on deep learning is proposed. The watermark is encrypted by double random phase encryption and the encrypted image is embedded in the host image. Then the physical relationship between watermark image and watermarked host image is used to train an improved neural network. Using the model of FC-DenseNets-BC can reconstruct the watermark image. In the traditional optical watermarking technology, the quality of watermarked host image and decrypted watermark image depends on the selection of the embedding strength. However, using deep learning to reconstruct the watermark image can get rid of this dependency. The simulation results show that the method can reconstruct high-quality watermark image with peak signal-to-noise ratio over 35dB even when the embedding strength is as low as 0.05. It also has a certain generalization, safety and ability to resist noise and shear. The feasibility and efficiency of this method are further verified by the experiment of optical system.