Imaging through scattering media with high fidelity is still one of the main challenges in imaging analysis of deep biological tissues and distant astronomical observations. The computational imaging method based on deep learning has made significant progress in reconstruction quality and other aspects. However, when the scattering media in the actual system is unstable and the structure of objects is complex, and the obtained scattering dataset for training is limited, the pure data-driven method cannot realize efficient reconstruction. An efficient imaging method was proposed in reconstructing complex objects through unknown thin scattering media with different statistical properties, which was based on the effective combination of the speckle correlation theory and the powerful data mining and mapping capabilities. More information had been unearthed with the redundancy of the speckles and had been fully used with the neural network. This method obtained high-quality recovery of complex objects with complex scattering scenes and the training set is limited. This approach can promote the applications of physics-aware learning in practical scattering scenes.