The real-time correction algorithm is proposed, which adjusts the integration time adaptively according to the current scene information in the process of correction. Firstly, the background images are stored at different integration time and the integration time is adjusted by the specific thresholds. Then the corresponding background image is subtracted from the output of IRPPA and the result are taken as the input of neural network. So it can make up the deficiency of neural network, and reduce the non-uniformity caused by the change of integration time. Experiments of real IRFPA videos show that the proposed algorithm insures high quality of image, also has the advantages of integration time adjusting adaptively and real-time correction.