Resolution Enhancement of Still Images and Video

The need to increase the resolution of a still image or video arises frequently in digital cameras/camcorder, security/surveillance systems, medical imaging, aerial/satellite imaging, scanning and printing devices, and high-definition TV monitors. However, due to the hardware limitation, we usually obtain low resolution images or video sequences. Therefore, the aim of this paper is to reconstruct a high resolution image/video from available low-resolution images/frames. In thesis, we address three studies in the area of resolution enhancement.
First, we propose an improved image super-resolution system based on iterative back projection (IBP). The system consists of three major modifications, the subpixel registration, key frame selection and blur estimation. Compared to the traditional method, the proposed system achieves better image quality in the situation of unknown blur and save time cost by choosing the minimum low-resolution frames.
Next, we study resolution enhancement in the spatial domain. As for image resolution enhancement, our approach is to recover high resolution of the least-zoomed image when given a sequence of images with different zoom factors of the same static scene. Other zoomed images are warped to align with the least zoomed image and discontinuity of the boundary region is blended. On the other hand, we combine high resolution still images and a low-resolution video sequence of the same scene to enhance the resolution of video. The proposed method gives the satisfactory visualization and significant quality improvement.
Finally, we study resolution enhancement in the compressed domain. In the discrete cosine transform (DCT) domain, we arbitrarily resize an image in terms of block DCT coefficients, process the edge enhancement by linearly high-pass filtering and remove block artifacts by weighting DCT coefficients. Compared to the bilinear interpolation, the output image is visually sharper and gives significant improvements in peak signal-to-noise ratio (PSNR). In the discrete wavelet transform (DWT) domain, the image to be arbitrarily resized is taken as the low frequency subband with appropriate scale. The enhanced image avoids blockiness and blur.