Research And Application Of Vision And Video Key Techniques Oriented To Scientific Instrument Network Lab

With the development of science and technology, the scientific research and technical development are more and more dependent on the advanced scientific instruments. The government pays more attention to the research conditions now, consequently, universities and research institutes purchase so many scientific instruments, and the quantity increases every year. Nevertheless, there is still lack of instruments when compared with the requirement from the development of college teaching and scientific research. The problem with the instrument usage can be summed up to three aspects: unbalanced distribution, the low utilization efficiency and short of experimental teaching resources. The construction of scientific instrument network lab, composed of laboratory comprehensive management and remote experiment with scientific instrument, can apply networked management to instruments and promote the usage efficiency of instrument resource. Also students and researchers can take part in the analytical experiment remotely which is convenient for the studies and researches. There are two key problems need to solve in the network lab, the first one is the method of work status monitor based on computer vision and the other is how to compress the sample video effectively. This thesis gave the advanced studies to the two problems. On the foundation of the location of status indicator, the scientific instrument work state is monitored by introducing the multi-feature fusion into the status detection with computer vision. According to the motion principle and features of sample video, the optimizations were made to the local motion estimation (LME) and the global motion estimation (GME). The main research contents are as follows:(1) The detection and location of status indicator based on computer visionTo monitor work status of scientific instrument automatically with computer vision, the first thing is the detection and location of the status indicator. Most of the status indicators are circular contour, so how to detect and localize the circular object was studied. By synthesizing the strongpoint of geometric characteristics and random sampling, an efficient method of detecting circles, called Semi-Random Detection (SRD) based on Right Triangles Inscribed in a Circle (RTIC), was presented. Above all, an Array Space based on the Positions of Valid Pixels (ASPVP) was constructed. And then four points searching method was used to search the corresponding right triangles. After employing a few specific and effective tools such as radius constraint, error distance judgment, and data merging etc., the method of deleting circles was applied to get the real circles. Some composite images with different levels of noises and some real images with status indicators in the scientific instrument had been taken to test the performance. Experimental results show that the proposed algorithm is accurate, fast, and reliable and meets the requirements of the detection of circular status indicator. As a universal algorithm, SRD is also suit for the usual circle detection.(2) The work status detection based on the image multi-feature fusionA series of system disturbances exit in the vision detection of the work status of scientific instrument, therefore, it is not feasible to detect the state with single feature. Multi-feature fusion is introduced into the status detection with computer vision. Because of the independent and complementary characteristic among the features, such as feature point location, template matching, hue matching and intensity matching, a series and parallel combination structure model (SPCSM) was proposed to build models for the fusion of the multi-feature. With fuzzy confidence of each feature as model input, the state of work indicator of scientific instrument was decided. Application experiments in electron probe (NO.EMX-SM7) show that our algorithm can execute monitoring task accurately, reliably on line in changeable environment. By the way, the idea of multi-feature fusion detection can be extended to other state detection fields.(3) Local motion estimation optimization with sample videoThe dynamic sample image sequence, formed during finding the positions of checking samples and scanning images when analyzing samples, is named sample video. When using the conventional motion estimation algorithms to sample video coding, the invalid search points will be searched redundantly and some valid points will be missed. The motion principle of sample video was studied, and as a consequence, there is a boundary scope during the motion of the sample video. According to the result, an Unsymmetrical motion estimation method with Boundary Constraint (UBC) was proposed. Firstly, on the base of original search scope, the motion performance of sample video from the instrument was measured and recorded with the"one meter one parameter"method. According to the statistical results, a motion boundary was set to narrow the search scope. And then, under the boundary constraint, an unsymmetrical search model was brought out to avoid the redundant search with invalid points and the miss from valid points. The different sample videos from electron probe and electron microscope were employed to test the performances of UBC. It proved that the UBC compresses the motion estimation time by 40% compared with the UMHexagonS method. And in terms of coding quality, UBC got the similar PSNR with UMHexagonS.(4) Global motion estimation optimization with sample videoThe motion of sample video takes on global, isotropic feature. Three single motion models were given for the three motion modes of sample video respectively, and then a global motion model suit for sample video was synthesized with the three single models. According to the global motion feature, by synthesizing the strongpoint of pixel-based and macroblock-based global motion estimation, an efficient MacroBlock Pair vector based GME method (MBP-GME) was proposed. Combined with the filter equation of feature macroblock, Outlier macroblocks, instead of outlier pixels, were eliminated in advance on the basis of global consistency of the macroblock pair. And then, simplified pixel-based gradient descent iteration was performed to get the final global motion parameters using the pixels in the feature macroblocks. During the iteration, the outlier pixels were further eliminated from the feature macroblocks. Several test sequences, sample videos and nature video included, were used to test the algorithm performance. Experimental results show that the proposed algorithm is accurate, fast and efficient.