Different from the static characteristics of solid scattering media such as ground glass, the scattering effect of turbid media on light beams is reflected both in the space and time domains. Most traditional scattering imaging methods are inapplicable to dynamic turbid media. To address this issue, a deep learning-based method is proposed to reconstruct objects in the presence of turbid media. The imaging quality of the proposed neural network under the conditions of different turbid media and turbid media with different concentrations is studied. The generalization ability of the neural network is tested. The experimental results demonstrate that high-quality imaging is achieved by the proposed network. Moreover, the network shows strong generalization ability and robustness under the mixed training of speckle images of turbid media with different concentrations.