| As the most significant source in China’s energy supply, Coal plays a decisive role in ensuring China’s energy security, and will continue to support national economic development in the foreseeable future. However, the mechanized level of coal mining in china is still low, which lead to low productivity and security risks, so that coal mining safety has become increasingly important. Realizing that it is important to introduce modern science and technology to improve the level of coal mining management, China has been promoting no/less people working site of coal mining nationwide in these years, so that remote video surveillance system has been developed rapidly. By using remote video monitoring system, staff on the ground can monitor underground work sites in real-time, detecting security risks and preventing mine accident. So that mine video surveillance system has a higher demand for image quality. But there are some cases in coal mine production:deep underground work sites have limited lighting conditions; dusty air and air humidity have impact on image acquisition; monitoring equipment used in the majority of mine is outdated, and shoot from far distance, so the target object is not clear and the image quality is not high; image transmissed from a distance is easy to introduce noise. So the surveillance images are usually ambiguous and contain a lot of noise.Improving the quality of image can be start from two aspects:hardware and software. It is possible to get more accurate picture by improving the image sensor and other hardware devices, but the economic cost of large-scale deployment is too high, and the improvement of image quality is also limited. Software method, that is image super-resolution restoration technology, uses prior knowledge or complementary information within a sequence of images to restore the high frequency information of an image degenerated by collection, transmission and storage. Selecting appropriate image super-resolution method to improve the quality and visual effects of the coalmine surveillance image become a major subject to researchers.The research of this thesis mainly consists of several aspects as follows:Firstly, three popular image super-resolution technologies have been studied, analyzed and compared, that is interpolation-based, modeling-based and learning-based super-resolution method. Since the learning-based super-resolution algorithms take fully advantage of prior knowledge to save the image texture, and is relatively less sensitive to noise, we choose to do further research on learning-based super-resolution considering the characteristics of coal mine surveillance images.By using histogram matching method and fuzzy K-means clustering in learning-based super-resolution method, an improved neighbor embedding method is introduced in this thesis. Using a feature vector combined by first-order gradient and Gaussian character to represent image block, which can retain both high and low frequency information of the image better, and avoid the effect of noise to some extent. By learning the training set with fuzzy K-means clustering, this method can effectively avoid the edge blur, and achieve better preservation of image texture to provide more details. Comparative experiments show that the high resolution images restored by method proposed by this thesis are better than those restored by classical kernel regression methods and iterative back projection measured by root mean square error and peak signal to noise ratio, and it performs much better on image which contains more texture area. This method can reduce affect of the noise to some extent, reduce overall blur effectively, retain higher contrast, and retain texture area more clearly. Experimental results show that this method can reflect the coalmine working face more clearly and accurately to meet the needs of the coalmine surveillance system. |
PDF Full Text Request