Because the distance between the infrared imaging system and the target is usually far away, the background in the infrared image often contains clouds, clutter, and infrared correction caused by uneven undulation or stripes as well as all kinds of noise. At the same time, the small target of image in the pixel is less, and the lack of shape and texture information, the signal-to-noise ratio (SNR) is low, easily submerged in the complex background, which makes the detection and tracking of infrared dim and small target difficult. In order to improve the abilities of de-tection and tracking of infrared dim and small targets, it is necessary to effectively suppress the complex back-ground in infrared images. The background prediction is a valid complex background suppression method. By dividing the original image from the predicted background, a differential image is obtained, and the difference image can sufficiently suppress the complex background and the target is effectively preserved.At present, the commonly used background prediction algorithms include low-pass filtering, median filtering, morphological filtering, two-dimensional least mean square error filtering (TDLMS), mixed of Gaussian, back-ground prediction based on pixel estimation method, and so on. The above background prediction methods are effective for the background of stable or slowly changing, but they are ineffective for the large span background. In view of the shortcomings of the above methods, considering the ability of gradient inverse weighting filter has the advantages of good detail preserving and strong clutter resistance, it is introduced into the paper for background suppression. However, because its key parameters cannot be adjusted in real time according to local clutter, in order to enhance its adaptability, this paper proposes an improved gradient inverse weighting filtering algorithm through the establishment of clutter local correlation function. The use of background of the local statistical char-acteristics of adaptive filter parameters, can better adapt to the drastic change in the clutter, and improve the ability to suppress clutter suppression algorithm.For evaluation of background prediction result, three performance indices are used in this study, mean squared error (MSE), structural similarity (SSIM) and local signal-to-noise ratio gain (GSNR), to evaluate the effect of image background prediction. By comparing and analyzing these three indexes, MSE, SSIM and GSNR, it could be seen that the improved gradient inverse weighted filtering could effectively suppress the complex background of images, presenting a superior overall performance to other background suppression methods.