Memorization and Processing for Image Data Using Associative Memory

项目成果总结为以下三点：1。提出了一种基于固定模式概率赋值的连续色调图像二层表示方法。使用该方法，可以将记忆图像数据的大小减小到原始大小的四分之一。研究了如何利用联想记忆构建图像数据库（associative image database）的问题。(1)提出了一种利用不完整（退化）的记忆图像数据构建关联图像数据库的方法。该方法利用空间约简方法从给定的不完全记忆图像数据中获得正确的记忆信息。由于在实际应用中很难得到完整的记忆图像数据，因此这种构造方法是有用的。(2)提出了不完整输入键的最优关联图像数据库。众所周知，对于不完全输入键，T.Kohonen（1973）提出的Moore-Penrose型联想记忆极其敏感（不稳定），联想误差大得令人无法接受。所提出的最优关联图像数据库（关联存储器）既实现了稳定关联，又实现了最小误差关联。该方法解决了在实际操作中很难得到完整输入键的问题。(3)提出了一种实用的针对不完整输入键的关联图像数据库构建方法。该方法易于实现稳定的关联，但关联误差并不总是最小的。这种方法的优点是简单。提出了一种基于联想记忆的图像处理（降噪）方法。由于该方法基于两步关联过程，因此可以通过改变中间向量的维数来控制噪声衰减的能力。利用该方法，可以在关联图像数据库中同时进行记忆图像数据的信息检索和信息处理。

The results of the project are summarized in the following three points:1. A new bilevel representation method for continuous tone images using a probablistic assignment of fixed patterns is presented. Using the method, the size of the memorized image data can be reduced to quarter its original size.2. The problems of how to construct image database using associative memory (Associative Image Database) is considerd.(1) A method to construct an Associative Image Database from incomplete (degraded) memorized image data is proposed. In the method the correct memorized information is obtained from the given incomplete memorized image data using space reduction method. The construction method is useful, because it is very rare to get complete memorized image data in practice.(2) The optimal Associative Image Database for incomplete input key is proposed. It is known that for incomplete input key the Moore-Penrose type associative memory proposed by T.Kohonen (1973) is extremely sensitive (unstable) and its association error becomes unacceptably large. The proposed optimal Associative Image Database (associative memory) realizes both the stable association and the minimum error association. The proposed method is useful, because it is very rare to get complete input key in practice.(3) A practical construction method of Associative Image Database for incomplete input key is also proposed. In the method the stabel association can be easily done, but the association error is not always minimum. The merit of the method is its simplicity.3. An image processing (noise attenuation) method using associative memory is proposed. Since the method is based on the Two-Stepped Association Procedure, the ability of noise attenuation can be controled by changing the dimensions of the mediate vectors. Using the method, both information retrieval and information processing for memorized image data can be done simultaneously in the Associative Image Database.