Study On Perceptual Grouping Of Coal-mine Image Based On Attention Mechanism

The surveillance pattern of current coal-mine video surveillance systems is too simple, they are just a system for real-time information collection and transmission. In fact, as the number of camera is increasing, it's becoming more and more difficult to capture the abnormal status, that results in high omission factor, can't satisfy the coal-mine demand for safety, so it's a potential trouble for coal-mine safety production.To develop the coal-mine video surveillance system, it is necessary to apply the intelligent surveillance technique on coal-mine image and video processing. To achieve the purpose, first of all, we must detect objects out of the coal-mine videos and images. Perceptual grouping can form the object hypothesis with the least specific knowledge, and reduce the computation complexity of visual recognition. This dissertation introduces perceptual grouping of visual information processing system to detect the coal-mine objects.But the application of perceptual grouping algorithm is restricted by some factors, such as environment condition, convergence speed, and adaption. Aiming at the three problems, this dissertation is mainly organized by three parts: (1) Extraction of grouping seed for coal-mine uneven light images; (2) Interactive closed boundary grouping based on double-weight graph; (3) Perceptual grouping algorithm of coal-mine complex image based on bottom-up attention mechanism and perceptual grouping algorithm of coal-mine complex image based on up-bottom attention mechanism.(1) Grouping seed extraction, which will greatly reduce the difficulty of following perceptual grouping steps, is one important process of perceptual grouping. But the environment conditions result in uneven lighting phenomenon which disturbs the accuracy of grouping seed extraction of perceptual grouping. In order to unlock the restriction of environment conditions, this dissertation puts forward the grouping seed extraction algorithm for coal-mine uneven light images. Firstly, we show the edge model of uneven lighting image, and then based on nonlinear visual perception characteristic, we introduce two edge detection algorithms for uneven lighting image, at last, we explain how to transform these coarse unsmooth and intersecting edges to smooth grouping seeds. The experiments show that grouping seed extraction algorithm based on non-linear visual characteristic reduces the redundant edges at bright regions and loss of edges at dark regions.(2) The current perceptual grouping algorithms are mostly aiming at general application. At the coal-mine field, maybe there are many objects which fulfill the grouping laws, so we must run the perceptual grouping algorithm many times to focus on the final needed object, this process costs more computation time, and will not definitely find the wanted result. This dissertation introduces an interactive closed contour grouping algorithm based on double-weight graph, we analyze the similarities and differences between interactive and non-interactive perceptual grouping algorithms. The experiments show that: the computation complexity difference between the interactive and the non-interactive algorithms is small, but convergence speed of interactive algorithm is at least 3 times faster than that of non-interactive algorithm;(3) If the interactive process of interactive perceptual grouping algorithm completely depends on manual selection, we must ask for the professional surveillance worker to do it, this action reduces the flexibility of algorithm. Attention mechanism simulates the human's visual process, directly locates around the most interesting object. This dissertation puts forward two attention mechanisms to subtitude the manual selection: one is based on bottom-up attention mechanism, the other is based on up- bottom attention mechanism. The experiments show that: the computation model of bottom-up attention mechanism run more than 4 times per second, but it has position deviation phenomenon which is about 40%. The perceptual grouping algorithm with specific knowledge based on up-bottom attention mechanism can select the global cue. Cue error ratio and seed error ratio are both low, when detecting only one seed, they both are zero, when below 12 seeds, they both are under the 40%.We have designed a platform implemented by Matlab and C++ for the experiment. The tasks of the platform include: grouping seed extraction, computation of global and local cues, solution of object function, realization of two kinds of attention mechanisms.The dissertation has 47 figures, 4 tables and 193 references.