Stacked convolutional auto-encoders for single space target image blind deconvolution

The blind restoration of a scene distorted by atmospheric turbulence remains a challenging problem for space target video surveillance. As there are multiple factors coupling the degradation of space target images, traditional methods based on a single, simplified image blind restoration model have difficulty achieving the desired results. In this paper, a new convolutional auto-encoder deep neural network is proposed for modeling the degradation and restoration process of spatial target images. The whole network consists of two parts, convolution and deconvolution, which are used to achieve the purpose of degraded feature learning and blind image restoration, respectively. Simulation image training data are constructed by a series of 3D space target models and combined with the turbulence multi-factor degradation model. The neural network model is then trained with this data. Contrast experiments are conducted using the simulation image data and real image data, and the results show that the proposed method is robust to noise and the reconstructed images have clearer edge details. The output images also have better continuous consistency and superior visual effects.