A generative adversarial net framework （GAN） based on the encoding-decoding architecture was proposed and used for computational optical imaging through scattering media. Training data set and testing data sets were generated from the original images.The GAN network proposed was trained with data sets generated by phase maps of original images and had the ability of mapping relationship between the speckle pattern and the original image. By experimentally analyzing the quality of scattering imaging with different data sets and under different scattering conditions， it is found that the GAN network has good generalization ability and could be used to realize the computational optical imaging through scattering media. Furthermore， the imaging quality with the three improved loss functions proposed were analyzed and compared. Experimental results showed that the imaging quality was effectively improved using the loss function with the structural similarity function. It is of great significance for computational optical imaging through scattering media by the proposed method.