To solve the problems of edge blurring and slow convergence speed faced when a traditional neural network method is used in the non-uniformity correction of an infrared focal plane array, the traditional neural network algorithm is improved by introducing the local gradient characteristics of images. By using the weighting function constructed with local gradient similarity information to weight a region, the image edge information can be preserved. In iterative computation, an adaptive weighting factor with gradient amplitude is added to the error loss function, and the adaptive step size associated with the gradient amplitude is introduced to replace the traditional fixed step size. Thus, the correction effect and convergence speed of the algorithm are further improved. Then, the performance curve of the algorithm and its correction result are analyzed. The result shows that the improved neural network correction algorithm has achieved better non-uniformity correction effect than the traditional algorithm. Its correction error is less than that of the traditional method. The object to effectively suppress edge blur and improve convergence speed is realized.