Infrared blind-pixel compensation algorithm based on generative adversarial networks and Poisson image blending

Infrared focal plane array is affected by nonuniformity, which leads to the generation of infrared image blind pixels and reduces the image quality. This paper puts forward an infrared blind-pixel compensation algorithm based on generative adversarial networks and Poisson image blending (GAN–PIB), which implements the greyscale prediction of the original image blind pixels through the combination of the pre-trained adversarial network and the blind-pixel compensation function and in the form of generating new images, breaking through the traditional thinking of the existing compensation algorithm based on interpolation and filtering. Firstly, the blind-pixel compensation network is constructed based on the generative adversarial networks. Through training, the model learns the image features of infrared blind pixels and achieves a better compensation effect for the blind-pixel image data sets which are used for training. Secondly, the blind-pixel detection of the test blind-pixel images is carried out to generate a binary mask matrix, and the constructed blind-pixel compensation loss function is combined to create the fake image. Finally, by fitting the original blind-pixel image and the generated image with the Poisson image blending algorithm, the compensation precision is improved through iteration, and finally, the compensation for the infrared blind pixels is completed. The experimental results show that GAN–PIB algorithm has strong adaptability to isolated blind pixels and clustered blind pixels, and compared with traditional algorithms, the compensated images have higher intelligibility and texture details.