Estimation of Similarity of Image Data Using Associative Memory and Its Applications

(1)提出了一种新的联想记忆模型，实现了退化输入图像数据的稳定联想和最小误差联想。该模型是在图像复原模型的基础上构建的。在Moore-Penrose广义逆联想记忆中，图像数据以矩阵的形式分布式存储。当结合小规模数据库构建大比例尺图像数据库时，需要对存储的矩阵进行合并。利用正交化投影的性质，给出了一种合并矩阵的方法。通过理论分析，给出了相似性估计能力与存储的原型图像数据数量之间的关系。(3)提出了Hopfield网络中能量函数的一种正交变换方法。在对网络进行编程时，能量函数由线性…定义目标函数和约束的更多组合。由于组合会造成干扰，解决方案将缺乏正确性或有效性。正交变换方法通过在另一个基上重新定义函数来改善这种缺点。验证了该方法对于图像数据的模糊聚类是可行的。(4)在传统的图像数据库中，为了估计两幅图像之间的相似性，需要从两幅图像中提取一些特征。选取了传统图像数据库中常用的典型特征，给出了这些特征的一些性质。(5)提出了一种简单的全彩图像数据压缩方法，大大节省了图像所需的比特数。该方法基于多值颜色抖动方法和传统的流行度算法。在该方法中，由于抖动的影响和随后适当的颜色选择，很容易获得颜色再现和空间分辨率，并且还抑制了强烈的轮廓。较少

The results of the project are summarized in the following five points:(1) A new associative memory model which realizes both the stable and minimum error association for degraded input image data is proposed. The model is constructed based on the image restoration model. The performance of the model is also analyzed theoretically.(2) In the Moore-Penrose generalized inverse associative memory, image data are memorized distributively in a matrix form. When a large scale image database is constructed by combining small scale databases, it is necessary to consolidate the memorized matrices. Using the property of orthonormal projection, a method to consolidate the matrices is introduced. From a theoretical analysis, we show the relation between the estimation ability of similarity and the number of memorized prototype image data.(3) An orthogonal transformation method of energy function in Hopfield network is proposed. In programming a network, the energy functions are defined by a linear … More combination of an objective function and constraints. Since the combination causes interference, the solution would lack correctness or validity. The orthogonal transformation method improves this kind of badness by redefining the function in another basis. We verify the method can work for a Fuzzy Clustering of image data.(4) In conventional image database, in order to estimate the similarity of two images, some features are extracted from them. Typical features which are often used in conventional image database are selected, and some properties of the features are given.(5) A simple method of full color image data compression technique that significantly economizes on the number of bits required for an image is proposed. The proposed method is based on the multi-valued color dithering method and the conventional Popularity Algorithm. In the method, due to the effect of dither and the ensuing adequate color selection, both color reproduction and spatial resolution is easily obtained, and strong contouring is also suppressed. Less