Beyond Camera Limits:Image Enhancement by Miso Image Super Resolution
【摘要】：The recent rapid development in image capturing devices provides many opportunities to people to get a better quality camera available in a market. However, due to their high prices many people didn't go for it and they still want to use their existing digital cameras. On the other side demand for High Resolution (HR) images is the cry of the day same as more and more bandwidth are required day by day for different bandwidth-hungry applications in communication area. Limited image resolution due to various degradations factors such as noise and blur leads to a number of problems in different multimedia applications such as object recognition in video surveillance. This fact brings out the demand for study different image processing tasks such as image enhancement and image restoration that aim to improve efficiently the interpretability of visual information lies in images for us. However, they didn't provide desired results after certain level of improvement. Super Resolution (SR) technology is a real science and engineering rather than fiction utilized to combat limited image resolution problem and promise to produce a desired HR output from a sequence of low resolution (LR) images, which forms the core issue discussed in this dissertation. Our work mainly focuses on construction of a Multi Input Single Output (MISO) system for SR. The MISO system has an input of multiple images, which can be taken by a video camera or still image camera and the output is a single image with higher resolution than the input images. The aim of generating a HR image is to uncover the details of the scene and increasing the number of pixels as well. The basis for MISO super resolution reconstruction is provided by multi-frame analysis while the quality improvement is caused by fractional-pixel displacements existed between multiple input images. Due to all this, SR allows to overcome the inherent limitations of the imaging system without the need for additional hardware modification. That's the reason that today it has a booming market demand.
This thesis addresses different aspects of MISO image SR. We explore various schemes in different processing domains to produce better HR image and to improve the performance of existing techniques as well.
In the first contribution, we perform well-known grid analysis for comparing the performance of various existing MISO image SR techniques. We considered different existing techniques for this purpose and taken different valuable factors into account to choose a better technique among them.
In the second contribution, we formulate an efficient algorithm based on recently introduced curvelets and used the absolute values of kurtosis as an input. Simulation results show that proposed algorithm provides significant PSNR gain and outperform existing transform-based technique.
The third contribution of the thesis addresses use of statistics to develop a simple pre-processing strategy for image SR reconstruction that exploit the optimal' selection of LR inputs and generate a HR image of same quality as produced by utilizing all the input LR images. Simulation results validate our proposed strategy using different kind of images.
The fourth contribution deals with implementation of an integrated SR reconstruction algorithm based on interpolation of cropped LR frames extracted from a low quality video surveillance sequence to perform effective license plate recognition. The proposed algorithm provides an alternative to modern costly surveillance cameras.
Finally, the novelty of last contribution is investigation of a relationship between LR input images and the resultant HR image for existing MISO SR techniques. We also address one of the most important quality assessment issues for MISO reconstruction-based image SR techniques in this contribution as well.