Research On Key Techniques Of Super-resolution Reconstruction Of Aerial Images

Spatial resolution of the image is a key technical indicator of aerial imaging system, which directly determines the effect of aerial reconnaissance. Aerial imaging process is affected by many factors such as external disturbances, motion blur, flight attitude changes, and cloud occlusion. Therefore the actual image resolution inevitably becomes lower, which brings difficulties to the subsequent image processing and analysis. Limited to manufacturing technology and cost, enhancing spatial resolution by improving the sensor size or reducing the pixel size of the imaging device will be costly. If pixel size is too small the signal-to-noise ratio(SNR) also drops, and deteriorates image quality. Therefore, study on how to use super resolution reconstruction technology to break the inherent limitation of imaging system, to restore the original appearance of the image as far as possible or to further improve the image resolution and definition is significant and indispensable.This dissertation discusses the basic theory and the development of dynamic image super resolution reconstruction technology. Various aspects of image super resolution reconstruction are studied systematically. Focusing on the main factor of aerial imaging environment degradation, super resolution reconstruction method for single frame motion blurred images, global motion image sequence, local motion image sequences are studied and implemented.The main innovations and research results of this dissertation are as follows:1. For motion blurred image, a novel technique that performs motion deblurring and super-resolution jointly from one single blurry image is presented. This method firstly constructs a multiscale image pyramid. In the kernel estimation step, the statistical characteristics of natural images with gradient sparsity are used to construct the L0 norm regularization constraint function. The accurate motion blur kernel is estimated by the alternate iteration method. Secondly,ATV regularization term based on L1 norm is used to constrain the cost function in super resolution reconstruction step. It relaxes the gradient domain constraints, and overcomes the constraint that the L0 norm is too harsh hence makes the image too smooth. By the improved method, natural reconstruction results are obtained. Finally, a method based on guided filtering is used for post-processing to further enhance the quality of the reconstructed image and suppress ringing and noise.2. Image sequence registration and super resolution reconstruction algorithm are deep researched and validated by some cases. In a global motion model, an improved BRISK feature image registration algorithm is proposed. The original BRISK algorithm neglects corner distribution information and single matching strategy leads to a high rate of false match. To address this problem, firstly, we built a continuous scale space based on the original BRISK algorithm. After that, adaptive corner extraction threshold were selected by the image saliency map, to obtain the local extreme value and visual saliency feature points. Finally, the matching of binary features is carried out via FLANN algorithm and RANSAC is employed to eliminate wrong matches. These efforts provide more reliable registration parameters for the reconstruction stage. In the reconstruction step, the influence of norm selection on the reconstruction results is analyzed in detail for super-resolution reconstruction of regularization cost function. And the dual L1 norm which can preserves the image edge and improves the quality of the reconstructed result is chosen for super resolution reconstruction.3. For local motion sequence, a multi-frame super resolution framework combined with single frame enhancement is proposed. Firstly, we use the super-resolution convolutional neural network technique for feature-extraction and enhancement. Secondly, relative displacements of the high-resolution candidates are accurately estimated by a robust edge-preserving optical flow algorithm which applies a bilateral weight map based on spatial and color similarity to reduce the image ambiguity. Finally, a high-resolution image with more details is obtained as a weighted combination of high-resolution candidates for each pixel based on confidences map of high-resolution candidates. IBP algorithm is used to reduce the reconstruction error. This method combines the advantages of single frame and multi frame algorithm, and effectively solves the problem of multi-frame super resolution reconstruction in complex motion patterns.4. In order to improve the spatial resolution of aerial image, an image super resolution framework based on the polyphase components(PPCs) reconstruction algorithm is proposed. Though the assumption of diversity sampling, this method adopts a fundamentally different approach, in which the low-resolution frames is used as the basis and the reference frame as the reference subpolyphase component of the high resolution image for recovering the polyphase components of the high resolution image. The polyphase components, which fuse the low resolution frames with the complementary details, are obtained by computing their expansion coefficients in terms of the basis using the available sub-polyphase components and then inversely transforming them into a high resolution image. Furthermore, an improved steering kernel regression algorithm is proposed to help restore the fusion image with mild blur and random noise. The steering kernel regression function adaptively refined by the local region context and structures. Thus, this new algorithm not only effectively combines denoising and deblurring together, but also preserves the edge information. Our framework improves the computational efficiency via avoiding accurately registration and complicated iterative computation. It is of great significance for improving the quality of aerial image at the present stage.In conclusion, this dissertation analyzes many related techniques involving aerial image super resolution reconstruction. Key problems including motion blur, global motion, and local motion in complex aerial imaging process are discussed, and it has gotten primary achievements. All these researches not only provide a theoretical basis for the further research, but also improve the spatial resolution and the detail definition of aerial image.