Abstract: In the image description method, only the image is taken as the input when the information is input. In the end-to-end training process, the internal parameter transformation is difficult to obtain, which may cause errors. In order to further reduce the uncertainty of image description, knowledge enhancement method is applied in the image description task. Subject informations are input at the input port to define the scope of the image description. A new two-channel image description architecture is proposed. The architecture includes two parts: theme channel and image channel. The theme channel extracts semantic information and uses semantic information as theme information to enhance the knowledge of image information. Image channel is of the classic image description task function. Features are encoded extracted by the fast regional neural network in the two channels, and feature screening is carried out by the attention mechanism, while the prediction information is decoded by the long and short term memory network. Finally, a long and short term memory network is used to synthesize the information of the two channels to enhance the knowledge of the main channel and generate the description of the image channel. This method is tested on the data sets Flickr and MS COCO. Compared with the general image description method, the results show that the accuracy is improved.