| With the increasing importance of space exploration,and the intensification of space competition among countries.More and more countries pay attention to space object monitoring and target recognition.The work of monitoring and recognition depends on the space object image captured by optical lens.Therefore,it is a great significance to obtain high-quality space target images for space surveillance and space attack defense confrontation.However,due to the large number of various rays in the space environment,the optical lenses and detectors on the satellite are not enough to support high-resolution imaging,so it is difficult to obtain high-quality space object images.Under the background that the hardware conditions are difficult to improve,the image super-resolution reconstruction technology can play a role,which using algorithms to restore the image from low resolution to high resolution.However,it is a typical ill-conditioned inverse problem to infer high resolution image from low resolution image.We hope that the images generated by super-resolution algorithms are accurate(low distortion)and realistic(high perception).but now,most of the current super-resolution algorithms only optimize one of the directions,because it is very difficult to achieve the best in these two directions at the same time.Generally,the low distortion image is often low in perception,which does not conform to the human eye cognition,while highperception images have serious image distortion.To solve the above problems,this paper proposes a super-resolution reconstruction method of space object image based on wavelet transform and deep learning.It aims to achieve that the space object image reconstructed super-resolution reconstruction has high perceptual quality while reducing distortion.Specifically,the main contributions of this paper are as follows:(1)A new super-resolution model of spatial target image based on convolutional neural network is proposed.This model combines the spatial domain and the wavelet domain,The network input is a low-resolution image,and the output is sub-bands of the hight-resolution wavelet coefficients.Using the characteristics of perfect reconstruction of wavelet transform,these sub-bands of wavelet coefficients are recombined to obtain images with both higher perceptual quality and lower distortion.(2)In order to balance the distortion and perception of the reconstructed image of the space object,this paper designs the perception network and the objective network,and uses wavelet transform to separate the low-frequency information related to the objective quality and the high-frequency information related to the perceived quality from the image.Then,optimize them according to the different goal.This ”divide and conquer” strategy achieves a good balance between the perceived quality and objective quality of super-resolution images,and achieves better visual effect.(3)Compared with the traditional wavelet transform-based method,our method directly takes the low-resolution image as input,learns and completes the wavelet transform in the neural network,Instead of,enlarge the low-resolution image to the high-resolution image through bicubic up-sampling interpolation and perform wavelet transform,Then the transformed wavelet coefficients are used as the network input.So the computational complexity of neural network is greatly reduced.Moreover,in the field of space object image,due to its complex imaging environment,the use of bicubic interpolation to enlarge the image will cause the reconstructed image to have artifacts.Therefore,this method which does not rely on bicubic up sampling interpolation can obtain better performance for space object image reconstruction.(4)The traditional neural network is mostly one-to-one learning mapping relationship,while the neural network model in this paper uses multi-channel output to realize one-to-four mapping relationship.Comparing the performance of multi-channel output and single channel output in wavelet domain,this paper finds that no obvious performance difference between them,but the number of convolution layers required for multi-channel output is only 1/n(n is the number of channels)of single channel,which further improves the learning efficiency of the network.Experimental results show that the space object image reconstructed by our method has a good balance between distortion and perception,and has good performance and reconstruction quality.Moreover,quantitative indicators,the peak signal noise ratio(PSNR)of the results obtained by using our method proposed in this paper is 0.86 d B higher than the super-resolution method based on the generative adversarial network. |
PDF Full Text Request