Multimode fiber is a thick scattering medium. When the target image is projected onto the multimode optical fiber, multimode coupling will occur, thereby generating speckle images at the output of the fiber. In this work, multimode optical fiber imaging is restored based on deep learning, and the distortion of thick scattering media imaging is solved. DenseUnet is used and the speckle image is used as the input of the model for reconstructing the target image. The DenseUnet model employs a fusion mechanism to deepen the network depth, thus, improving the reconstruction accuracy and realizing good robustness. The experimental results reveal that DenseUnet can be used to reconstruct speckle images produced by multimode optical fibers with different lengths.