For the application of conventional infrared image non-uniformity correction method, when the integration time is adjusted, the image gray level may change accordingly, an infrared image nonuniformity correction method adapted to adjustment of integration time was proposed. In this method, the black-body calibration images with different integration time and temperature, and the corresponding theoretical infrared radiation were integrated into a whole database, with the loss function of neural network and error reverse transfer mechanism, the correction coefficients in the model were learned during the training. The trained network correct non-uniformity can ensure the stable output of images during the real-time adjustment of infrared camera integration time, it is significant for post infrared image processing. And it is proved that the training of this network does not need enormous calibration datum. In order to deal with the problem of infrared detector response drift, an online learning correction coefficient method is proposed.