Research On Key Technologies Of Multiscale Image Super-Resolution Reconstruction

For a given imaging system, we can improve the resolution of the image by some direct methods, such as improving the lens, changing the circuit and optical design, increasing the digit size in ADC, reducing the quantization error and so on. But these direct methods are limited by the high economic cost and craft level, so they cannot be used widely. Super-resolution reconstruction of the image refers to the technology that with the existing imaging conditions, we use software to eliminate the distortion and degradation of the image which is produced by the imaging system, reconstruct the high frequency information of the image and reconstruct a clear picture with high spatial resolution. Therefore, super-resolution becomes a hotspot in digital image processing and computer vision in recent years and this technology is of great practical significance in high-resolution imaging systems and infrared imaging systems. We choose face image as our study objects, because compare with natural images, face images have more fine and complex features and super-resolution for face image is more challenging and has certain practical significance. Face image super-resolution can be applied in many fields, such as face recognition, facial expression analysis and so on.In this paper, we study multiscale image super-resolution reconstruction algorithm and realize the learning-based super-resolution algorithms using face images. We propose the multiscale super-resolution reconstruction algorithm and also do experimental analysis. Details are as follows:1. We propose a multisacle super-resolution algorithm for face image based on steerable pyramid. In this algorithm, steerable pyramid is used to capture the spatial distribution of low-level local features in face images, and then these features are combined with ptramid-like parent structure and local optimal mathching algorithm based on parent structure to predict the best prior. After that, the prior is integrated into the Bayesian maximum a posteriori probability framework to get the high-resolution face image. At last, the optimal high-resolution face image is obtained by a global linear smoothing operator. Experiments show that the oriented facial features are recovered well and the edge of the image is smoother. The high-resolution face image is of better visual effects.2. Base on the learning-based super-resolution framework, we propose a multisacle super-resolution algorithm for face image based on Contourlet pyramid. In this algorithm, we use Contourlet pyramid to learn the low-level local features in face images and combine Contourlet coefficient characteristics and gradient characteristics together as the feature representation of the image. Finally, we use the local optimal macthing based on image pathches to obtain the high-resolution face images. Experiments show that the algorithm can achieve better experimental results and enhance the details in the reconstructed face image.