Constraint Satisfying Image Processing Using Top-down Information

为了对图像进行识别，我们需要基于构造推理的知识处理，因为图像中没有足够的图像识别信息。自顶向下的过程被认为是控制图像处理区域或图像处理算法参数的控制系统。在本研究中，我们提出了“约束满足图像处理”的概念，这是对自顶向下信息图像处理算法的推广，并实现了这一概念。在以前的图像识别系统中，对自下而上的图像处理结果进行对象假设生成等知识处理。然而，这种方法有点像“平滑模糊的结果”，这是图像识别系统不能实际使用的主要事实。因此，我们构建了“约束满足图像处理”的概念，将需要什么样的对象，或者需要什么样的结果的信息提供给图像处理系统。为了实现“约束满足图像处理”，我们讨论了约束的表示。因此，我们意识到可以通过特征值层次表示的约束来控制图像处理。活动轮廓模型（蛇）是基于能量最小化原理的正则化方法的一种轮廓检测方法。我们开发了一种机制，通过提供表示必须检测哪种轮廓的信息来控制蛇检测所需的轮廓。为了在各种情况下实现“约束满足图像处理”，我们做了用约束的信念因子来控制过程的经验。

In order to recognize an image, we need knowledge processing based on constructive inference, because there are not enough information for image recognition in an image. Top-down process, which is used to verify candidates, were recognized as control system which controls the processing area of images or parameters of image processing algorithm. In this research, we propose the concept of "Constraint Satisfying Image Processing" which is generalization of image processing algorithm which uses top-down information, and we implemented this concept.1. In the former image recognition system, knowledge processing like object hypothesis generation, were applied to the results of bottom-up image processing. However, this approach is something like "smoothing over vague results" and this is the main fact that the image recognition system can not be used practically. Therefore, we constructed the concept of "Constraint Satisfying Image Processing" which gives the information of what kind of object is needed, or what kind of result is needed, to the image processing system.2. We discussed about the representation of constraints in order to realize "Constraint Satisfying Image Processing". As a result, we realized that the image processing can be controlled by the constraints of feature value level representation.3. Active Contour Model (snakes) is a contour detection method which is an application of regularization method that is based on energy minimization principle. We developped the mechanism which controls the snakes to detect needed contour by providing information which represents what kind of contour must be detected.4. We have done the experience of controlling the process by belief factor of constraints in order to realize "Constraint Satisfying Image Processing" in various situation.