Image segmentation is the most difficult problem in visual computation, whose difficulties lie in its intrinsic uncertainties. This lead to many corresponding approaches based on different applications and physical constraints. It is urgent that we should have a sound framework in analyzing all these algorithms, where we are expected to catch the key issues of them, put forward efficient solutions and, moreover, facilitate our current visual computation both in theories and applications. Based the consideration mentioned above, the main issues we have tried to tackled contains: ◆ In the shadows of the fruits of visual computation theories, We proposed a new framework analyzing the current segmentation algorithms, where we first introduced the concept of "data-driven" and "model-driven". We deem that regularazation theories and bayes-based statistical inference are intrinsically same in the sense of the optimization of energy function when tackling ill-posed problems but the latter one has a wider meaning when come to describing uncertainties in visual computations. ◆ We discussed the segmentation models based on MRF-MAP framework; and we deduced Gibbs probability distribution through maximum entropy principle in statistical physics, which verified the unified visual computation theories in the sense of the optimization of energy function. ◆ We proposed the multi-resolution representation of image to tackle the difficulties of segmentation. In additions, we dig into the pyramids theories and proved qualitatively that it is expected to get good parameters estimation results using pyramid representation; And We got the unified way to generate pyramid representations under the framework of the optimizations of energy function and introduced a promising idea called nonlinear pyramid. ◆ We dig deeply into the algorithms of parameter estimation and proposed an algorithm integrating EM and MDL to get the parameters and the number of regions intended. We also introduced a novel algorithm called Adaptive Edge Propagation to optimizing the MRF-MAP function, Which can deal with the inherent problems in former algorithms proposed and get satisfactory results. Comparative experiments results to others algorithms showed the superiority of our theory.
修改评论