Research On Key Technologies Of Multi-frame Image Super-resolution Reconstruction

Spatial resolution is typically an important measure for image quality, the higher the resolution is, the more details of the original scene the image can provide. Due to the physical limitations and high cost of optical devices, there is urgent need to develop a new method breaking through imaging system limits to improve resolution, which is the research purpose of super-resolution image reconstruction technique in the field of computer vision and image processing. Recently, super-resolution image reconstruction technique is widely applied in various imaging fields such as medical imaging, satellite imaging, video monitoring, remote sensing and etc. Therefore it has high theoretical significance and application value.In this paper, multi-frame image super-resolution reconstruction techniques are studied. The purpose is to fuse the relevant information between a sequence of low-resolution(LR), blurred and noisy images (or a low resolution video sequence) with the same scene, and eventually to generate a high-resolution (HR) deblurred and denoised image (or a high-resolution video sequence). Specific contents are as follows:(1) From the perspective of the static reconstruction, multi-frame image super-resolution reconstruction technique based on Gaussian pyramid optical flow (GPOF) registration and L1 norm is proposed. The motion estimation model in the method employs an idea of optical flow sub-pixel registration based on Gaussian pyramid hierarchical structure, which not only accelerates the implementation of the algorithm but also achieves the sub-pixel accuracy. In the reconstruction process, this method is based on robust L1 norm both in data fidelity term and regularization term. The bilateral total variation (BTV) priori model is employed as a regularization term which not only decreases the computation but also keeps the image edge. Finally, we employ median"shift and add"idea to initialize the HR image value in the objective function optimization iteration equation, when the motions between LR frames are pure translations and the blur is space invariant. Experiments show that the method can remove outliers efficiently, resulting in image with sharp edges.(2) Based on the idea of static super-resolution reconstruction, the paper proposed a dynamic reconstruction approach for the monochrome video sequences based on approximation of the Kalman filter (KF). For the case of translational motion and common space-invariant blur, the approach provides a recursive model and forward data fusion method according to the Kalman filter update equation for the input LR video sequence, which is so-called the dynamic shift-and-add method. By updating the mean-covariance pair in the update equation, the method generates an initial blurred HR video sequence (?)(t) using a causal mode. Finally, we propose a theory combining the MAP estimate with BTV prior to deblur and interpolate the initial HR video sequence for obtaining the final high-resolution video sequence. Experiment indicates that the reconstruction process is of efficient storage and low computational cost, and better results can be achieved.