机载红外点目标探测系统在搭载飞机飞行中探测系统的环境参数会发生变化，导致通过传统地面标定方法获取的非均匀性校正参数的准确性有所降低，故有必要进行机上基于场景的非均匀性校正。本文提出了一种基于帧间配准的机上非均匀性校正算法，首先对图像进行预处理，滤除探测器坏点影响，然后用两帧邻近图像计算互功率谱，求出互相关函数，确定配准位移。两帧连续图像完成配准后，通过误差函数最小化来实现校正参数的更新，最后对整个图像序列进行上述迭代计算，获取最终校正参数。本文模拟了一组非均匀性场景图像序列作为实验图像序列，通过实验分析，提出了帧间图像变化（平移、旋转、缩放）对本算法校正效果的影响，然后采用两个具有代表性的算法与本文提出的算法分别对该图像序列进行处理，并从图像质量和收敛速度两方面比较算法性能。结果表明：与其他两种算法相比，本文提出的算法非均匀性校正效果较好，峰值信噪比提高了20 dB以上，结构相似性则突破了0.99。本文提出的算法虽然比较复杂，但校正参数收敛速度较快，易于在硬件平台上实现，具有一定的工程应用前景。

During flight, changes in environmental parameters affect the accuracy of non-uniformity correction of airborne infrared point-target detection systems to some extent. Therefore, it is necessary to perform on-board scene-based non-uniformity correction. In this paper, we propose an algorithm based on inter-frame registration to achieve on-board non-uniformity correction. It first preprocesses the images to filter out dead pixels, then calculates a cross-power spectrum with two adjacent frames and determines their registration displacement according to a correlation function that is calculated from the cross-power spectrum. After registering an image, corrective parameters are updated by minimizing the error through a function. The corrective parameters are finally obtained after the above calculations are performed. In an experimental comparison, we simulate a set of non-uniform scene image sequences as the experimental image sequence. This experiment first identified the influence of changes between frames (including displacement, rotation and scaling) on the accuracy of non-uniformity correction. It then used two representative algorithms and the proposed algorithm to process the image series, and compared the performance of the algorithms from the perspective of image quality and convergence speed. The results show that the proposed algorithm has better non-uniformity corrective performance compared with the two other methods. The PSNR increased by over 20 dB, and the SSIM exceeded 0.99. The proposed algorithm has higher complexity, but its convergence speed is much faster, and it is easy to be implemented on hardware platforms, which gives the algorithm possible applications in engineering.